Simcenter Nastran 2019.

1
PARALLEL PROCESSING GUIDE

1
 Proprietary & Restricted Rights Notice

© 2019 Siemens Product Lifecycle Management Software Inc. All Rights Reserved.
This software and related documentation are proprietary to Siemens Product Lifecycle
Management Software Inc.
NASTRAN is a registered trademark of the National Aeronautics and Space
Administration. Simcenter Nastran is an enhanced proprietary version developed and
maintained by Siemens Product Lifecycle Management Software Inc.
MSC is a registered trademark of MSC.Software Corporation. MSC.Nastran and
MSC.Patran are trademarks of MSC.Software Corporation.
All other trademarks are the property of their respective owners.

TAUCS Copyright and License

TAUCS Version 2.0, November 29, 2001. Copyright (c) 2001, 2002, 2003 by Sivan
Toledo, Tel-Aviv University, [email protected]. All Rights Reserved.
TAUCS License:
Your use or distribution of TAUCS or any derivative code implies that you agree to this
License.
THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
EXPRESSED OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
Permission is hereby granted to use or copy this program, provided that the Copyright,
this License, and the Availability of the original version is retained on all copies. User
documentation of any code that uses this code or any derivative code must cite the
Copyright, this License, the Availability note, and "Used by permission." If this code or
any derivative code is accessible from within MATLAB, then typing "help taucs" must
cite the Copyright, and "type taucs" must also cite this License and the Availability note.
Permission to modify the code and to distribute modified code is granted, provided the
Copyright, this License, and the Availability note are retained, and a notice that the code
was modified is included. This software is provided to you free of charge.
Availability (TAUCS)
As of version 2.1, the code is distributed in 4 formats: zip and tarred-gzipped (tgz), with
or without binaries for external libraries. The bundled external libraries should allow you
to build the test programs on Linux, Windows, and MacOS X without installing
additional software. It is recommended that you download the full distributions, and then
perhaps replace the bundled libraries by higher performance ones (e.g., with a BLAS
library that is specifically optimized for your machine). If you want to conserve
bandwidth and you want to install the required libraries yourself, download the lean
distributions. The zip and tgz files are identical, except that on Linux, Unix, and MacOS,
unpacking the tgz file ensures that the configure script is marked as executable (unpack
with tar xvpf), otherwise you will have to change its permissions manually.

2
Table of Contents
Table of Contents ................................................................................................................ 3
About this Book .................................................................................................................. 5
Chapter 1 Introduction and Fundamentals ........................................................................ 6
1.1 Parallel Computing in Simcenter nastran ................................................................. 6
1.2 Keywords in Simcenter nastran Parallel Computing ................................................ 7
1.3 Getting Started .......................................................................................................... 8
1.4 Parallel machine architectures .................................................................................. 9
1.5 Parallelism in Simcenter nastran ............................................................................. 10
1.6 Parallel Solution Methods in Simcenter nastran ..................................................... 11
1.7 Expectation from Simcenter nastran Parallel Solutions.......................................... 12
Chapter 2 Running Parallel Solutions ............................................................................. 13
2.1 Computational Methods at a Glance ....................................................................... 14
2.2 Running SMP Jobs.................................................................................................. 18
2.3 Environment Setup for DMP .................................................................................. 19
2.4 Running DMP Jobs ................................................................................................. 20
Chapter 3 Methods for Linear Static Analysis ................................................................ 22
3.1 Geometric Domain Static Analysis (GDSTAT) ..................................................... 22
3.2 Load Domain Static Analysis (LDSTAT) .............................................................. 24
3.3 Recommendations for the Method Selection .......................................................... 26
Chapter 4 Methods for Normal Modes Analysis ............................................................ 27
4.1 Geometric Domain Normal Modes Analysis (GDMODES) .................................. 27
4.2 Frequency Domain Normal Modes Analysis (FDMODES) ................................... 31
4.3 Hierarchic Domain Normal Modes Analysis (HDMODES) .................................. 36
4.4 Recursive Domain Normal Modes Analysis (RDMODES) ................................... 41
4.5 Recommendations for Method Selection ................................................................ 45
Chapter 5 Methods for Response Analysis and Optimization ........................................ 47
5.1 Frequency Domain Frequency Response Analysis (FDFREQR) ........................... 47
5.2 DMODES + FDFREQR for SOL 111 .................................................................... 50
5.3 DMODES + Serial transient calculation for SOL 112............................................ 51
5.4 DMODES + serial optimization process for SOL 200 ........................................... 53
Chapter 6 Methods for Multi-step Nonlinear Solution ................................................... 57
6.1 Geometric Domain Static analysis (GDSTAT) in SOL 401 ................................... 57
6.2 Normal modes analysis subcase in SOL 401 .......................................................... 58
6.3 Cyclic modes or Fourier modes analysis subcase in SOL 401 ............................... 61
6.4 DMP for a Multi-step example ............................................................................... 65
Chapter 7 Performance Study ......................................................................................... 70
7.1 Performance of Parallel Processing ........................................................................ 70
7.2 Industrial Case Study 1 ........................................................................................... 70
7.3 Industrial Case Study 2 ........................................................................................... 74
Chapter 8 Installation and Configuration of DMP .......................................................... 77
8.1 Overview ................................................................................................................. 77
8.2 Requirements .......................................................................................................... 77
8.3 Windows Single Host Instructions.......................................................................... 81
8.4 Windows Multiple Host Instructions (True Cluster) .............................................. 83

3
 8.5 Determining Hosts .................................................................................................. 94
8.6 Managing Host-Database Directory Assignments .................................................. 96
8.7 Managing Files........................................................................................................ 97
8.8 Performance Issues ................................................................................................. 97
8.9 Overview of Running a DMP job ........................................................................... 98
Chapter 9 GPGPU computing....................................................................................... 100
9.1 Overview ............................................................................................................... 100
9.2 GPU Keyword Input Summary............................................................................. 100
9.3 Modal frequency response (FRRD1 module) Example ........................................ 101
Appendix A ..................................................................................................................... 104
SEQP Style DMP Solutions ........................................................................................ 104
Appendix B ..................................................................................................................... 111
GPGPU Library Installation ....................................................................................... 111
References ....................................................................................................................... 112

4
About this Book

SIMCENTER NASTRAN is a general-purpose finite element program which solves a

wide variety of engineering problems. The SIMCENTER NASTRAN Parallel Processing
Guide is intended to help you choose among the different parallel processing and
computational methods, and ultimately increase the performance of analysis by reducing
CPU time, memory and disk space requirements.

This main material in this book covers the parallel processing methods for the linear
static, normal modes, direct frequency response, modal frequency response, modal
transient response, and design optimization.

This book is composed of 9 chapters:

1. Introduction and fundamentals

2. Running parallel Simcenter nastran solutions
3. DMP computational methods for linear static analysis
4. DMP computational methods for normal modes analysis
5. DMP computational methods for response analysis and optimization
6. DMP computational methods for multi-step nonlinear analysis
7. Performance study
8. Installation and Configuration of Distributed Memory Parallel (DMP)
9. GPGPU Computing

To effectively use this book, it is important for you to be familiar with the basic structure
of Simcenter nastran. For more information about the mathematical foundation, refer to
Simcenter nastran Numerical Methods User’s Guide.

This book will continue to be revised to reflect future enhancements to Simcenter nastran
parallel processing methods. Changes and additions are encouraged, and can be
communicate through Siemens PLM technical support http://support.ugs.com/.

5
Chapter 1 Introduction and Fundamentals

In today’s Finite Element Analysis (FEA) applications, the volume of data is increasing
rapidly, the physics being modeled is increasingly complex, and the time allocated to
development is under increasing pressure. Meanwhile, most computational advances
have come in the realms of increased core counts, increased SIMD processing (i.e.
vectorization) capabilities, and larger and deeper memory hierarchies. In order to delve
into larger and more complex problems, FEA software must better leverage these
computational advances. Clearly, a focus on optimizing processes and data flows for
serial computational performance will not suffice.

Parallel computing is an approach that uses multiple computers, processors or cores

working together on a common task. Each processor works on a section of the problem
and exchanges information with other processors.

1.1 Parallel Computing in Simcenter nastran

With parallel computing, Simcenter nastran provides simultaneous use of multiple

processors on one or more machines to decrease solution times. The benefits of parallel
processing include:

 Reduced solution time on large problems

 Decreased turnaround time for each solution; allowing you to analyze more designs
under more conditions in a shorter amount of time
 Reduced hardware requirements by utilizing smaller and less expensive machines

The following are basic parallelism concepts in Simcenter nastran.

Types of parallelism. There are basically two types – data parallel and task parallel. In
data parallel, each processor performs the same task on different data. In task parallel,
each processor performs a different task. Simcenter nastran, like most applications,
utilizes both data and task parallelism in the course of most analyses.

Parallel computer architectures. The control mechanism of parallel computing in

Simcenter nastran is based on Multiple Instruction, Multiple Data (MIMD) which refers
to a parallel execution model in which each processor is essentially acting independently.
The processors work on their own data with their own instructions. Tasks executed by
different processors can start or finish at different times.

Parallel programming models. Simcenter nastran supports shared memory (or threads)
and message passing. These parallel programming models exit as an abstraction above
hardware and memory architectures. In hybrid parallel models, two or more parallel
programming models are combined.

6
Message passing is a widely used communication model in programming parallel
processor machines for distributed memory parallel (DMP) processing. You can access
the communication hardware only through the message passing library. Message passing
libraries like Message Passing Interface (MPI) or Parallel Virtual Machine (PVM) send
and receive data from one process to another. The two processes can be running on the
same machine, or two entirely separate ones. In particular, MPI is a standard message
passing library that has been efficiently implemented on a variety of platforms.

MPI is portable, efficient, expressive, and provides thread safety. In addition,

 Point-to-point communications handle data transmission between any two
processes in a communicator.
 Collective communications handle simultaneous communication between all
processes in a communicator.

Simcenter nastran offers the ability to run certain solution sequences in parallel using a
Message Passing Interface (MPI), an industry-wide standard library for C and Fortran
message-passing programs. MPI programs can be run on SMP computers, DMP
computers, and a cluster of computers supported by the MPI package.

Further information on the MPI standard is available online at the MPI forum website:
http://www.mpi-forum.org.

1.2 Keywords in Simcenter nastran Parallel Computing

The following keywords are used in parallel Simcenter nastran.

Execution keywords: dmp, hosts, slaveout, parallel, numseg, nclust, nrec, dstat.

Performance keywords: memory, scratch, sdirectory, sscr, buffsize, rdscale.

In the appendix, gpart is described as an execution keyword, and it is typically not

required. More keywords are described in the Simcenter nastran Installation and
Operation Guide.

7
1.3 Getting Started

The following steps give you a basic idea how to use parallel processing for the
desired analysis in Simcenter nastran.

 Step 1. Know your machine architectures.

SMP or DMP machine (see 1.4 Parallel machine architectures), memory size, etc.
 Step 2. Choose parallelism scheme.
SMP or DMP scheme (1.5 Parallelism in ). If SMP only, refer to 2.2 Running
SMP Jobs. Most solution sequences, and the RDMODES method (), support SMP.
If DMP, refer to 2.4 Running DMP Jobs. In many cases, a combination of SMP
and DMP is possible.
 Step 3. Select and run an appropriate method for the desired analysis.
Refer to 2.1 Computational Methods at a Glance.
 Step 4. Tune the performance.
Achieving the best performance may depend on the machine architecture. Refer
to 2.4.4 Other Considerations and for general guidelines.

8
1.4 Parallel machine architectures
Hardware can be divided roughly into two categories.

Shared Memory Parallel (SMP) Machines. A SMP machine is defined as a single

machine with multiple processors that share common memory and I/O as illustrated in
Figure 1.1. Multiple processors can operate independently, but changes in a memory
location affected by one processor are visible to all other processors. The primary
disadvantage is the lack of scalability between memory and processors.

CPU CPU CPU CPU

MEMORY

DISK

DISK

DISK

Figure 1.1 SMP Machine architecture

Distributed Memory Parallel (DMP) Machines. A Distributed Memory Parallel (DMP)

machine uses multiple machines or clusters with one or more processors communicating
over a network, or multiprocessors with multiple I/O channels. Figure 1.2 shows a typical
architecture of a DMP machine. Each machine has its own memory and one or more
disks. In DMP, data is private to each node and it is necessary to exchange data across
different nodes. Therefore, the programmer must decide which data is to be sent/received
to/from which node. The main advantage is that memory (and local disk) is scalable with
multiple processors.

Network

CPU CPU CPU CPU

MEMORY MEMORY MEMORY MEMORY

DISK DISK DISK DISK

Figure 1.2 DMP Machine architecture

9
1.5 Parallelism in Simcenter nastran

Simcenter nastran supports both shared memory parallel (SMP) and distributed memory
parallel (DMP) processing.

In Simcenter nastran, SMP is used only for lower level operations such as matrix
decomposition and matrix multiplication for all solution sequences. Therefore, as long as
suitable hardware is available, all solutions can utilize SMP processing.

DMP processing is based on domain decomposition of geometric, frequency, or load

domains. DMP methods achieve their solution speed by dividing the FE model into
smaller pieces to be solved simultaneously. This division can be performed with respect
to geometry or frequency range individually, or both at the same time. Although each
processor is working on its own partition of the geometry or frequency range, it
communicates with the others to share information. Once the solution is complete, the
results are merged, creating a single result file.

Most of the discussions and examples in this book focus on the DMP computational
methods and solution methods.

The differences between DMP and SMP are listed in the table below.

Feature Hardware Basis for Parallelism Software

Environment Mechanism
High-performance Partitioning of finite Message Passing
DMP workstation, or cluster element model Interface (MPI)
thereof
Shared memory multi- Partitioning of matrix and OpenMP
SMP processor computer vector operations API (or pthreads)

10
1.6 Parallel Solution Methods in Simcenter nastran

Simcenter nastran offers options for using parallelism across several solution sequences,
for both static and dynamic analyses. As mentioned in the previous section, two parallel
models (SMP and DMP) are available.

In SMP, Simcenter nastran parallelizes the decomposition of and arithmetic operations

(e.g. multiplication) on matrices. Since every solution sequence involves at least matrix
multiplications, SMP can be activated in all solution sequences for all analyses as long as
the hardware supports SMP. An SMP RDMODES run is available for SOL 103 and SOL
111 that can significantly reduce regular run time if an approximate solution is acceptable.
See 2.2 Running SMP Jobs on how to activate SMP.

In contrast to SMP, which focuses on parallelizing computational modules (such DMCP,

MPYAD), DMP provides parallelism on the algorithm level, which can provide greater
speedup. They can be categorized into the following three methods:

 Domain Static Analysis (DSTAT) method (see chapter 3)

 Domain Normal Modes Analysis (DMODES) method (see chapter 4)
 Domain Frequency Response Analysis (DFREQR) method (see chapter 5)
 Harmonic modes analysis method (see chapter 6)

They are for linear and nonlinear static analysis, normal modes analysis, and frequency
response analysis, respectively. Modal frequency and transient responses require
computing modes for modal space. Therefore, DMODES can be applied during mode
computation. In design optimization, if it involves computing modes, DMODES also can
be activated. In summary, DMP computational methods support the following solution
methods.

 SOL 101 Linear statics (see chapter 3)

 SOL 103 Normal modes (see chapter 4)
 SOL 105 Buckling (see chapter 4)
 SOL 108 Direct frequency response (see chapter 5)
 SOL 111 Modal frequency response (see chapter 5)
 SOL 112 Modal transient response (see chapter 5)
 SOL 200 Design optimization (see chapter 5)
 SOL 401 Nonlinear statics; Fourier, cyclic, and normal modes (see chapter 6)

The DMP computational methods will be discussed in Chapters 3-6.

11
1.7 Expectation from Simcenter nastran Parallel Solutions

A Simcenter nastran parallel solution should be able to solve a large problem with
significantly less run time compared to a standard serial (single processor) run. Note that
(as in most parallel codes), speedup will be less than the number of processors, due to
MPI startup cost, communication overhead, and inherent limitations in the
parallelizability of the algorithms used. Furthermore, if the computation can already be
done on a single processor in minimal runtime, there will not be much opportunity for
improvement, especially for small problems (less than 10,000 DOFs).

12
Chapter 2 Running Parallel Solutions

A Simcenter nastran parallel job can be selected by the Nastran command with either
keywords or system cells. The Nastran command permits the keywords, dmp and
parallel, to request DMP and SMP runs, respectively. Alternatively, system cells 231 and
107 can also be used for DMP and SMP runs.

It is strongly recommended to use keywords to request a parallel job. The following table
provides keywords and their descriptions as well as the system cells that can be used for
DMP and SMP runs.

Keyword System Cell Description

Default=0 (deselect DMP processing)
dmparallel Specifies the number of tasks for a DMP analysis
DMP 231
(or dmp) 1-512 processors are available in a DMP job
Default=0 (deselect SMP processing)
parallel Specifies the maximum number of CPUs selected
SMP 107
(or smp) for SMP processing
1-1023 processors are available in a SMP job

13
2.1 Computational Methods at a Glance
The SMP (shared memory parallel) computational methods are

Analysis Computational Submittal Suggested

Method Command Model Type
All Analysis:
None parallel=p All models
SOL xyz

One of the computational methods that is based on the domain decomposition but can run
in serial fashion (not DMP) or SMP is RDMODES, as an alternative to ACMS. It can
reduce the runtime significantly even on a single processor, compared to a conventional
single processor Lanczos run. Unlike other domain decomposition methods, RDMODES
does not require DMP.

Computational Submittal Suggested

Analysis
Method Command Model Type
All Analysis:
None parallel=p All models
SOL xyz
Normal modes
Large models, allow
calculation:
RDMODES parallel=p, nrec=n approximate
SOL 103
solution
SOL 401
Modal frequency RDMODES +
analysis: serial/SMP frequency parallel=p, nrec=n Large models
SOL 111 calculation

14
The DMP (distributed memory parallel) computational methods are:

Analysis Computational Submittal Suggested

Method Command Model Type
Linear static GDSTAT dmp=p Large models
analysis: Small models with large
SOL 101 LDSTAT dmp=p, dstat=1
number of loads
Nonlinear static
analysis: GDSTAT dmp=p Large models
SOL 401
Large models, small frequency
GDMODES dmp=p range request, also support
buckling (SOL 105)
Small models, large frequency
Normal modes FDMODES dmp=p, numseg=p
range request
analysis:
SOL103 Large models, large frequency
HDMODES dmp=p, nclust=c
range request
Large models, allow
RDMODES dmp=p, nrec=n
approximate solution
Direct frequency
response analysis: FDFREQR dmp=p All models
SOL 108

The DMP computational methods with application of normal modes computation are:

15
 Modal Additional
Submittal
Analysis Computational Computational
Command
Method method
Modal frequency
GD/FD/HD/RD- FDFREQR response Same as normal
analysis:
MODES calculation modes analysis
SOL 111
Modal transient
GD/FD/HD/RD- Serial transient Same as normal
analysis:
MODES calculation modes analysis
SOL 112
Design
GD/FD/HD/RD- Serial optimization Same as normal
optimization:
MODES process modes analysis
SOL 200
Normal modes
GD/FD/RD- Same as normal
analysis subcase N/A
MODES modes analysis
SOL 401
Cyclic or Fourier
Serial/SMP Lanczos Harmonics distributed dmp =p
modes analysis
as even as possible
subcase
Serial/SMP RDMODES among processors dmp =p, nrec=n
SOL 401

Note that RDMODES deactivates the sparse eigenvector recovery option in SOL 200.
The suggested problem type for a particular method is the same as in SOL 103.

Figure 2.1 Overview of Simcenter nastran DMP solution tasks

In Figure 2.1, the method selection for SOL 101 is described in Section 3.3. The method
selection for DMODES is described in Section 4.5. Overview is summarized as follows.

16
 Linear Statics Normal Modes

Yes

NREC>0 RDMODES
No

Yes Yes

NUMSEG>0 FDMODES
DSTAT=1
No

No (default)
Yes

NCLUST>0 HDMODES
No

GDSTAT LDSTAT GDMODES

17
2.2 Running SMP Jobs
The following example illustrates how to run a Simcenter nastran job named
‘example.dat’ in an SMP environment with p processors.

In SOL 103 and SOL 111, one also can specify RDMODES for large problems as
follows: 4

Where n is the number of external components. Refer to RDMODES (section 4.4) for
how to choose the number of components.

18
2.3 Environment Setup for DMP

System Prerequisites
Simcenter nastran supports the following systems:

X86_64 Linux and Windows-64

The detailed requirements of hardware and software are described in Chapter 7 -

Installation and Configuration of DMP, and in the Simcenter nastran Installation and
Operation Guide.

Message Passing Prerequisites

Simcenter nastran uses a Message Passing Interface (MPI) to manage a DMP task. Each
compute node must be able to access its local data. It is also necessary to communicate
between compute nodes. For these purposes, each local node (host) must have:

 Simcenter nastran installed properly. Simcenter nastran must be properly

installed on all the hosts listed by the "hosts" keyword or in the ‘host.list’ file.
 MPI program available. The MPI program start command must be available in
the path of the local host. (for example, "mpirun”)
 Input data file (including all bulk data and include files) must be accessible on
the local host.
 Do NOT assign output file names in data file.
 "r-" commands available and configured properly. such as rsh, rcp, and rlogin
to communicate between nodes.
 “scp” and “ssh” are supported on Linux. Need to put “s.rcp=scp” and
“s.rsh=ssh” and set the environment variable MPI_REMSH=ssh. These “s.”
commands can be in the command line or in the nastran.ini file.

More details are described in Chapter 7 - Installation and Configuration of Distributed

Memory Parallel (DMP).

19
2.4 Running DMP Jobs

2.4.1 Quick Start

When running a Simcenter nastran DMP job, it is necessary to specify a computational
method for desired solution sequence and a host list of the compute nodes. The Nastran
keywords DMP and HOSTS are required in a DMP run. See section 2.1 for keywords to
activate a particular DMP computational method.

[computational method] [host list]

Here is an example to run GDMODES with 4 processors for a SOL 103 example, called
‘example.dat’, on a four-node DMP machine. Assume these nodes are named node1,
node2, node3 and node4.

Notes:
 The "master" node is the first computer named by the "hosts" keywords, and
"slave" nodes are the remaining systems.
 The nastran keywords are processed in both the local and master/slave system.
 It is strongly recommended to use the nastran keyword "slaveout=yes" to print out
analysis procedures from all processors. With "slaveout=yes", the f04, .f06,
and .log files contain the outputs of master and slave processors.

2.4.2 I/O Enhancement Consideration

The performance of a Simcenter nastran parallel job is very much dependent on the CPU,
memory system, and I/O system performance. A DMP job is extremely sensitive to I/O
system performance, since each task independently accesses the I/O system. Specifically,
the performance of the disk subsystem that contains the permanent and SCRATCH
DBSets can have a significant impact on Simcenter nastran performance. This impact is
even greater if multiple tasks are using the same file system.

The scratch directory can be on a global or local file system. Siemens PLM recommends
that the "sdirectory" be local to each node, if possible.
The following example illustrates how to run ‘example.dat’ in parallel with four
processors with "hosts", and "sdirectory" local to each node.

20
Note that MIO may be used to reduce the overall run time by using some available
computer memory for I/O buffering. The nastran keyword ‘mio_cachesize’ is used to
control such memory size. See the Simcenter nastran Installation and Operation Guide.

2.4.3 Running DMP Jobs with Restart Option

There are two steps to run jobs using the restart option: cold start and restart. The first
step, called the cold start, generates the database that will be used in the second step. The
restart step runs related jobs with the given database. Note that the first step can be run
either in serial or DMP, while the restart step can be run only in serial.

Here is an example where the cold start is a SOL 103 job running DMP on 2 processors,
and restart is a SOL 103 example ‘exampler.dat’ on node1. Assume that the 2 processors
are named node1 and node2.

Note that the database name in cold start is example. Since the cold start is a DMP run,
the database saved on the master node (to be used for the restart) will be labeled as
example.t0. The database example.t1 from the slave node is not required for restart.
Hence for restart jobs, the given database is example.t0.

2.4.4 Other Considerations

 Memory (memory) – additional memory usually benefits DCMP; the memory

usage high water generally occurs in the partitioning module (GPARTN or
GPARTNS) for large models. Over-allocation should be avoided in order to have
better performance. Generally, allocated memory should be less than 80% of
available physical memory (as a rule of thumb, 50% or less for best I/O
performance on Linux systems).
 Scratch (sscr) – Adequate scratch space should be provided.
 SMP and MIO could affect the performance of a DMP job.
 Restarts are supported for a SOL 103 cold start in DMP and serial 103 and 111
restart.
 Contact conditions can be included in SMP and RDMODES runs. See Recursive
Domain Normal Modes Analysis (RDMODES).
 Contact conditions cannot be included in GDMODES, FDMODES, or
HDMODES runs.

21
Chapter 3 Methods for Linear Static Analysis

3.1 Geometric Domain Static Analysis (GDSTAT)

The GDSTAT provides an efficient parallel solution for the linear static analysis of large
models. The static solution is performed on the l-set. The l-set is identical to the a-set if
there is no rigid body support (r-set). The mathematical expression for the static
equilibrium of the finite element model can be expressed as:

K ll ul  Pl

where K ll is the global stiffness matrix, Pl is the load, and u l is the displacement.

The finite element model in the GDSTAT method is automatically partitioned into p
domains, where p is the number of processors. Figure 3.1 shows each domain that
contains the portion of the geometry (O1, O2, O3, or O4) plus the boundary t. Think of
geometric decomposition as an automated super-element approach. After the domain is
partitioned, each processor performs the linear static analysis only on its local domain
with boundary. GDSTAT allows a problem that is not possible to solve on one processor
to be solved in the DMP environment by reducing a significant amount of disk space and
memory.

O1 O2

O3 O4

t
Figure 3.1 Partitioning a finite element model for geometry domain static analysis
(GDSTAT)

22
In terms of partitioning, GDSTAT is available with both the GPARTN module and the
SEQP module. The GPARTN module is chosen by default with gpart=1 (the default
value). The module SEQP is activated by setting gpart=0. Note that the GPARTN-based
GDSTAT does not handle rigid body support (r-set). To avoid rigid body support,
constraints should be added using SPC rather than SUPORT cards.

The following example applies to the GPARTN module only. An example for the SEQP
module will be given in the appendix.

Example:

The static analysis is performed for a finite element model that has 2635 grid points for
200 CHEXA elements and 1210 CTETRA elements. GDSTAT is implemented in a DMP
run. The module GPARTN partitioned the global finite element model into four sub-
domains and a boundary. The following information in the .f06 describes the detailed
statistics of partitioning for the 1st subdomain.

Each processor performs static analysis with the corresponding local domain and the
boundary, and the master processor collects the results that you requested through
communications. These outputs are printed only on the master processor.

It is strongly recommended to use the nastran keyword "slaveout=yes" to print out

analysis procedures from all processors. With "slaveout=yes", the f04, .f06, and .log files
contain the outputs of master and slave processors in the following format.

23
 


















3.2 Load Domain Static Analysis (LDSTAT)

LDSTAT is useful when there are a large number of load cases in the linear static
analysis problem K ll ul  Pl . Instead of partitioning the finite element model, the load
matrix Pl is partitioned among the processors as evenly as possible, and the linear
solution is calculated within each of the respective processors for its own load cases. The
mathematical expression for the linear static analysis can be expressed as:

K ll u k  Pk , k  1, , p

where p is the number of processors. Once all processors finish their own linear solutions,
the master processor collects and forms the overall solution as

24
 ul  [u1 , u 2 , , u p ]

Note that each processor contains the full model, so that finite element model partitioning
is not required. LDSTAT is applied to a large number of loads with the same boundary
condition, i.e. only one stiffness matrix K ll is solved in LDSTAT.

Example:

For a model with 500 subcases, 125 subcases are assigned to each processor when four
processors are available. Without partitioning into domains, each processor performs the
static analysis with the partitioned load cases. The outputs are printed only on the master
processor after the master processor collects the requested output results from the slaves

25
3.3 Recommendations for the Method Selection

SOL 101 has two DMP methods for static analysis. It is recommended to use
GDSTAT if you have a very large model. LDSTAT is useful when there are large
numbers of load cases and a relatively small model.

The selection between GDSTAT and LDSTAT depends on the DSTAT keyword. The
method selection is described in Figure 3.2. The default value of DSTAT is 0.

SOL 101

FINITE ELEMENT MODEL

Yes

DSTAT=1

No (default)

GDSTAT LDSTAT

Figure 3.2 DMP Linear Static Analysis (SOL 101)

26
Chapter 4 Methods for Normal Modes Analysis

To analyze the dynamic behavior of structures, modal solution techniques are commonly
used. The essence of the modal solution is efficient calculation of the mechanical
system's free, undamped vibration. That is the eigenvalue analysis problem of:

( K aa  M aa )a  0

Here the eigenvalue  represents a natural frequency, and the eigenvector  is a free
vibration shape of the finite element model. The a subscript refers to the a-set, which is
the analysis partition of the finite element model. This is one of the most time consuming
computations of large scale global analyses in many applications such as the automobile
and the aerospace industries. This chapter describes available DMP computational
methods. These DMP methods are

 GDMODES Geometric domain normal modes analysis (see section 4.1)

 FDMODES Frequency domain normal modes analysis (see section 4.2)
 HDMODES Hierarchic domain normal modes analysis (see section 4.3)
 RDMODES Recursive domain normal modes analysis (see section 4.4)

The recommendations for method selection are also given in the section 4.5.

4.1 Geometric Domain Normal Modes Analysis (GDMODES)

GDMODES is executed by automatically subdividing the geometry obtained from the

finite element model. Such a subdivision of a finite element model is shown in Figure 4.1.
Here the o partition refers to the interior of the domains and the t partition is the common
boundary shared by the domains.

The geometric domain decomposition is mathematically represented with the following

reordering for the p partitioned domains.

 K oo
1
 M oo
1
K ot1  M ot1   o1 
  

2
K oo  M oo
2
K ot2  M ot2  o2 
       0
  
 K oop  M oop K otp  M otp  op 
 K T ,1  M T ,1 T ,2
 M oo
T ,2
 T,p
 M oo
T,p
K tt  M tt  t 
 oo oo K oo K oo

27
This partitioned eigenvalue problem may be solved by a special formulation of the
Lanczos method. Once each processor contains the portion of the local domain with
boundary, each processor computes its part of the global eigensolution and exchanges
boundary components with other processors.

Table 4-1 shows the eigensolution obtained from each processor, in which the global
eigensolutions are obtained by merging all distributed eigensolutions.

Notes:
 The formulation yields a computationally exact solution. The efficiency depends
on the relative size of the boundaries with respect to the interiors.
 The geometric domain decomposition is important to reduce the very large
problem sizes, but it does not affect the frequency spectrum.
 It is important to use the EIGRL card, not EIGR. In addition, the number of
processors p should be greater than 1. The value of the keyword "dmp" must be
an integer greater than or equal to 2, and a power of 2. For example, 2, 4, 8, 16,
etc. are valid.
frequency range

[ fmin fmax ]

O1 O2

O3 O4

t
Figure 4.1. Geometry domain partitioning for the normal mode analysis of a finite element model

Table 4-1 The distribution of eigensolution ( ,  ) with GDMODES parallel run, where
  {1 ,  2 , ,  p }T .

28
 partition 1 ( 1 ,  )

 
Geometry
partition i ( i ,  )
domain

 

partition p ( p ,  )

In terms of partitioning, GDMODES uses GPARTN by default (see appendix). The

module GPARTN performs both degree of freedom-based and grid-based partitioning, in
which p domains are created by an automatic partitioner from the connectivity graph of
the model. PARAM OLDSEQ can be used to specify the desired ordering method.

OLDSEQ Description
10 Metis with super-nodal and grid compressions
MLV with super-nodal and grid compressions
11
(default)
110 Metis with super-nodal compression
111 MLV with super-nodal compression
210 Metis with grid compression
211 MLV with grid compression

Notes:
1. If both super-nodal and grid compressions are selected (OLDSEQ=10 or 11), then
the GPARTN returns the coloring with smallest boundary.
2. The default OLDSEQ value is 11. System (294) = 1 prints additional diagnostic
information to the f06 file.
3. Note that in some cases, grid compression produces much smaller boundary size
than supernodal compression. As a result, the eigensolver (READ module) in
GDMODES can have a large run time difference.

Example:

A finite element plate model with 110 grid points and 100 CQUAD4 elements is
executed in parallel with SOL 103 analysis. The total number of degrees of freedom is
660. Four processors are used for the DMP run.

29
The partitioning statistics from the GPARTN module are shown as:

Note that the output is related to the degrees of freedom (DOF), not grid points, even
though the output uses GRIDS terminology instead of DOF. This example has 88 internal
degrees of freedom and 74 boundary degrees of freedom for domain 1. There are 91 and
74, 88 and 74, 91 and 74 degrees of freedom for the interior and boundary of domain 2, 3,
and 4, respectively.

The .f06 file of the master processor prints the summary of eigenvalue analysis.

However, with the gpart=1 option, the master processor does not broadcast the collected
eigenvalues and/or eigenvectors to the slave processors, so no eigensolutions are printed
in the.f06 file except the eigenvalue summary.

30


4.2 Frequency Domain Normal Modes Analysis (FDMODES)

31
In contrast to geometric domain partition, FDMODES is built on the frequency segment
approach. The frequency range of interest specified on the EIGRL entry is automatically
decomposed into multiple frequency segments—one for each processor. Table 4-2 shows
the frequency segment of each processor when s processors are available.

Table 4-2 Partition of frequency for FDMODES parallel run

Frequency Lower Upper

Processor
segment frequency frequency
1 1 f0 f1
2 2 f1 f2
   
j j fj-1 fj
   
s s fs-1 fs

Note that each processor contains the full model in the FDMODES computation, so that
the mode shapes in the individual frequency segments are independent of each other.
Figure 4.2 represents the schematic diagram: each processor solves the full model within
its frequency segment. The only communication needed is when gathering the results for
the master processor. Table 4-3 shows the eigensolutions obtained from each processor.

Notes:
 Although FDMODES reduces the frequency range for a process by decomposing
the frequency domain, it is still ineffective with respect to large problem size.
 Each processor contains the full model, so that finite element model partitioning
is not required. For best load balance, V1 and V2 of EIGRL should be specified
and ND omitted.
 In the case of repeat runs, the user can introduce the intermediate frequencies
directly on the EIGRL continuation card to enforce better load balancing than the
automatic generation.

32
 frequency range segment

[ f0 f1 ] [ f1 f2 ] .... [ fs-2 fs-1 ] [ fs-1 fs ]

( K aa  M aa )a  0

Figure 4.2 Frequency domain partitioning for the normal mode analysis of a whole finite element
model.

Table 4-3. The distribution of eigensolution ( ,  ),   {1 ,  2 , ,  s } and

  diag{1 ,  2 , ,  s } , with FDMODES parallel run

Frequency domain

Segment 1 … Segment j … Segment s

(1 , 1 ) … ( j ,  j ) … ( s ,  s )

Example:

This is the example used in GDMODES. FDMODES does not require domain
decomposition. The .f06 file shows the number of eigenvalues calculated in each
processor. Each processor computes 94, 61, 26, and 27 eigenvalues, respectively. The
total number of modes found is 208.

33
The master processor collects the eigenvalues and eigenvectors of slave processors, and
prints out the merged eigensolutions.

With "slaveout=yes", you can see the information about the eigenvalue problem on each
slave processor. It also shows the number of eigenvalues found on the slave processors.

34


35
4.3 Hierarchic Domain Normal Modes Analysis (HDMODES)

The HDMODES scheme simultaneously combines the two previously presented methods,
GDMODES and FDMODES. With this approach, a subset of processors or a cluster
solves the eigenvalue problem for the local geometry while communicating with other
subsets of processors or other clusters in order to consider the other frequency ranges.

The HDMODES computation is based on the processor assignment shown in Table 4-4,
where p * s processors are available. For each frequency segment, p processors are used
for GDMODES. In other words, s processors are used for FDMODES for each geometric
portion. For example, the ((j - 1) p+i)-th processor computes the eigenvalues of the j-th
frequency segment  j and the i-th geometric partition of the corresponding
eigenvectors  ij . The selection of the s and p value is problem dependent.

Note that HDMODES solves the eigenvalue problem computationally exactly, just as the
GDMODES and FDMODES methods do.

Table 4-4 Hierarchic domain decomposition concept

Frequency domain
Segment 1 … Segment j … Segment s
1 (j-1)*p+1 (s-1)*p+1
Partition 1 (1j ,  j )
(11 , 1 ) (1s ,  s )

   
Geometry i (j-1)*p+i (s-1)*p+i
Partition i ( ij ,  j )
domain (1i , 1 ) ( sj ,  s )

   
p j*p s*p
Partition p ( pj ,  j )
(1p , 1 ) ( sp ,  s )

The preferred hardware environment for HDMODES is a cluster of multiprocessor

workstations that is usually tied together by either a hardware switch or a network, as
illustrated in Fig. 4.3.

Assuming m workstations with n processors each, based on the scheme of Table 4-4, the
tasks of each column (the geometric partitions of a particular frequency segment) reside

36
on one workstation. The tasks of each row (the various frequency segments of a
particular geometry partition) are spread across the cluster, as shown in Fig. 4.4.

NETWORK

CPU CPU CPU CPU

M M M M
E E E E
M CPU M CPU M CPU M CPU
O O O O
CPU CPU CPU CPU
R R R R
Y CPU Y CPU Y CPU Y CPU

DISK DISK DISK DISK

Figure 4.3. A scheme of cluster of multi-processor workstations

P1, S1 P1, S2 ... P1, Sm

M M M M
E E ... E ... E ...
M P2, S1 M M M
O
... O ... O
Pi, Sj O ...
R R R R
Y Pn,S1 Y Pn,S2 Y ... Y Pn,Sm

DISK DISK DISK DISK

Figure 4.4. Mapping the hierarchic domain decomposition with n geometric domain partition and
m frequency partition (Pi: geometric domain partition and Sj: frequency segment) to a cluster of
multi-processor workstations.

The HDMODES solution sequence combines two existing techniques of DMP

processing: GDMODES and FDMODES. While the geometry partitions are solved
within a set of processors called a cluster, frequency segments are also solved in parallel
across multiple clusters. The advantage is a faster eigenvalue problem solution time for
very large models.

The keyword "dmp" defines the number of processors p, and the keyword "nclust"
defines the number of clusters c. The number of geometry partitioning g in a cluster does
not have to be defined explicitly. Note that 1<c<p and p=c*g. The number of clusters c
should be properly selected so that g is an integer greater than or equal to 2, and a power
of 2. For example, 2, 4, 8, 16, etc. are valid.

37
Example:

In the example below, HDMODES is executed with ‘dmp=4’ and ‘nclust=2’ keywords.
Among the four processors, the first processor is the master processor, and the other
processors are the slave processors.

Note that HDMODES defines the local master and local slave processors inside of each
cluster.

The partitioning statistics with GPARTN describe the number of degrees of freedom for
each cluster. In the example below, the first cluster has 223 degrees of freedom in the
interior of domain 1, and 50 degrees of freedom in the boundary. The second cluster has
the same local size as the first one.

In the .f06 file, the master processor prints the summary of the eigenvalue analysis and
the list of eigenvalues. Be careful when interpreting the "number of roots found"
information in the summary. This information concerns the cluster in which the master is
included. For example, in the model below, interpret the "number of roots found"
message as indicating that cluster 1 found 155 eigenvalues. The master processor lists all
208 eigenvalues that are merged from all of the local master processors. You should
determine the total number of modes from the list of eigenvalues.

38


Again, be cautious when interpreting the output from slave processors. The summary of
eigenvalue analysis for slave processors is confined to the corresponding processor. In
the list of eigenvalues on the slave processors, the mode number does not represent the
global mode number.

With gpart=1, the master process does not broadcast the collected output to slave
processors, so no eigenvalues are listed in the information from the slave processors.

39
40
4.4 Recursive Domain Normal Modes Analysis (RDMODES)

The Recursive Domain Normal Modes (RDMODES) analysis extends the DMP parallel
capability via substructuring technology for very large scale normal nodes analysis. It is
currently available in SOL 103 and 111. It also supports modal analysis in superelement
jobs. The RDMODES approach generally computes fewer modes with lower accuracy
compared to standard Lanczos approaches in order to gain performance.

RDMODES begins with partitioning the eigenvalue problem Kx  Mx into nrec
external partitions. The following represents a reordering for the nrec=4 with total 7
(=2*nrec-1) components for matrix K .

 K oo
4
K ot4, 2 K ot4,1 
 5 
 K oo K ot5, 2 K ot5,1 
 * * 2
K oo K ot2,1 
 
K  6
K oo K ot6,3 K ot6,1 
 7
K oo K ot7 ,3 K ot7 ,1 
 
 K ot3,1 
3
* * K oo
 * 1 
 * * * * * K oo 

(here the asterisks denote the transpose of the corresponding blocks in the upper
triangular portion, as the matrices are symmetric). Matrix M has the same structure.

This partitioned eigenvalue problem may be solved by a special formulation of the sub-
structuring method. All components are distributed evenly into processors. Each interior
eigensolution corresponding to its external partition is performed in serial, independent of
the others.

Contact conditions can be included in an RDMODES run. The input file should include a
static subcase with the contact conditions, and a consecutive normal modes subcase
which includes the STATSUB case control command. When you run with RDMODES
and contact conditions combined, an automatic static condensation is performed by
default during the static portion of the solution such that the contact iterations occur in a
reduced representation. As a result, performance gains occur in both the static and the
RDMODES portions. To run RDMODES without the static condensation, include the
parameter setting PARAM,RDCNT,NO.

The following is not supported when you combine RDMODES and contact conditions:
 Inertia relief, which is defined with the INREL parameter.
 The constraint mode method of enforced motion. The absolute method can be
selected instead with the system cell ENFMOTN.

41
RDMODES Sparse Eigenvector Recovery
In many instances, a user is only interested in the solutions at a few key locations instead
of all degrees of freedom, especially for large problems with millions of degrees of
freedom. In such cases, the sparse eigenvector recovery method can significantly reduce
the overall computation time and storage resource.

In RDMODES, the sparse eigenvector recovery option will be determined automatically

based on the user’s output request. If full eigenvectors are desired with only few output
requests, a user can deactivate sparse data recovery with PARAM,RDSPARSE,NO in the
BULK data.

RDMODES with the rdsparse option supports residual vectors (PARAM, RESVEC),
panel participation factors (PARAM, PANELMP), absolute displacement enforced
motion (sys422=1), and modal contributions. Note that PARAM,RESVINER is not
supported.

Note that the accelerated residual vector calculation with RDMODES takes advantage of
the rdsparse option, and is more efficient than the original one in terms of computational
time and I/O usage. The residual vectors with the accelerated calculation may differ
slightly from the original, which cannot be used in conjunction with rdsparse. If
necessary, the original resvec method may be requested by specifying PARAM,
RDRESVEC, NO in the bulk data. In this case, the rdsparse option will be disabled
automatically, which is likely to result in dramatically reduced performance.

RDMODES Eigensolver Selection

The RDQSOL parameter is available to request the eigensolver for the reduced
eigenvalue problem in the RDMODES method. The choices are LANC(ZOS), and
INCO(RE).
 The default, RDQSOL=LANC, selects the Lanczos method and is suitable for
most models.
 RDQSOL=INCO selects the in-core method, and is the best choice for dense
models.

Running RDMODES
RDMODES is activated by the Nastran keyword nrec. It can run in serial, SMP, and
DMP with optional keywords nclust=c and rdscale.

42
Notes:

1. p is an integer equal to a power of 2. It can be 1 (i.e. p=1).

2. The keyword nrec must be a positive number. The efficiency of the method requires
carefully choosing m, which should not be too small or too large. A nrec value with
power of 2 is required for better performance. For a large problem (>1M DOFs) nrec
equal to 128 or 256 would be a good choice.

Number of grids NREC value

1 ~ 5,000 4
5,000 ~ 40,000 8
15,000 ~ 80,000 16
30,000 ~ 150,000 32
60,000 ~ 300,000 64
120,000 ~ 600,000 128
250,000 ~1,200,000 256
> 1,200,000 512

This table does not suggest a unique nrec value for a given model. The best choice will
depend on the model and on the user’s machine configuration. After Simcenter nastran11,
the user can define NREC=1 and the software will automatically compute a reasonable
number of partitions based on the model size. The resulting number of partitions will be
reported in the .f06 file.

3. The optional keyword rdscale is used to increase the accuracy of the solution. In most
practical circumstances values in the range of 1.0 to 5.0 are acceptable. The trade off is
that the computational time increases with higher values of rdscale. The default value is
2.5.

4. PARAM,OLDSEQ in the input file can be used to specify the method of creating the
substructures. OLDSEQ is an existing parameter for geometric domain partitioning.

OLDSEQ Description
10 Metis with super-nodal and grid compressions
MLV with super-nodal and grid compressions
11
(default)
110 Metis with super-nodal compression
111 MLV with super-nodal compression
210 Metis with grid compression

43
 211 MLV with grid compression

Example:

In the example below, RDMODES is executed with ‘dmp=4’ and ‘nrec=8’ keywords.
Among the four processors, the first processor is the master processor, and the other
processors are the slave processors.

After the special formulation of the substructuring method, RDMODES calls 4-way
FDMODES for the reduced eigenvalue problem to obtain the global eigensolution. The
master processor collects the eigenvalues and eigenvectors of slave processors, and prints
out the merged eigensolution.
FREQUENCY DOMAIN PARALLEL LANCZOS METHOD

94 EIGENVALUES FOUND IN DISTRIBUTED SEGMENT # 1

61 EIGENVALUES FOUND IN DISTRIBUTED SEGMENT # 2

26 EIGENVALUES FOUND IN DISTRIBUTED SEGMENT # 3

27 EIGENVALUES FOUND IN DISTRIBUTED SEGMENT # 4

E I G E N V A L U E A N A L Y S I S S U M M A R Y (READ MODULE)

BLOCK SIZE USED ....................................... 7

NUMBER OF DECOMPOSITIONS .............. 4

NUMBER OF ROOTS FOUND .................... 208

NUMBER OF SOLVES REQUIRED ............ 43

REAL EIGENVALUES
(BEFORE AUGMENTATION OF RESIDUAL VECTORS)
MODE EXTRACTION EIGENVALUE RADIANS CYCLES GENERALIZED GENERALIZED
NO. ORDER MASS STIFFNESS
1 1 1.696475E+07 4.118829E+03 6.555319E+02 1.000000E+00 1.696475E+07
2 2 1.848272E+07 4.299153E+03 6.842314E+02 1.000000E+00 1.848272E+07

207 207 3.087263E+09 5.556315E+04 8.843149E+03 1.000000E+00 3.087263E+09

208 208 3.125765E+09 5.590854E+04 8.898121E+03 1.000000E+00 3.125765E+09

With "slaveout=yes", you can see the information about the eigenvalue problem on each
slave processor. It also shows the number of eigenvalues found on the slave processors.

44
4.5 Recommendations for Method Selection
It is recommended that you use FDMODES if you have a small model and a large
frequency range of interest. If you have a very large model and insufficient disk space,
use either GDMODES or HDMODES. Especially, HDMODES is recommended for a
cluster of multiprocessor workstations. RDMODES is recommended for large frequency
range, large models, and best performance when reduced accuracy is acceptable.

Fig. 4.5 illustrates the general guide for selecting a suitable DMP computational method
for normal modes analysis.

FDMODES
RDMODES
Frequency
HDMODES
range

GDMODES

FE model
size

Figure 4.5 General guideline for the selection of DMP computational methods

The selection of DMP computational methods introduced in this chapter is made

according to the keywords, NREC, NUMSEG, and NCLUST. The method selection is
described in Figure 4.6. Note the default value of NREC, NUMSEG and NCLUST is 0.

45
 SOL 103

FINITE ELEMENT
MODEL

Yes

NREC>0 RDMODES
No

Yes

NUMSEG>0 FDMODES
No

Yes

NCLUST>0 HDMODES
No

GDMODES

Figure 4.6 DMP normal mode analysis (SOL 103)

46
Chapter 5 Methods for Response Analysis and
Optimization

5.1 Frequency Domain Frequency Response Analysis

(FDFREQR)

The excitation frequencies, as specified on the FREQi entries, are split among the
processors as evenly as possible, and the responses for the partitioned excitation
frequencies are then calculated within each of the respective processors as shown in.
Fig.5.1. The whole range of excitation frequencies is partitioned into many sub-intervals
such as [Fi Fi+1]. Note that each processor contains the full model.

(Direct or Modal)
Frequency Response Problem

[ F0 F1] [ F1 F2] .... [ Fs-2 Fs-1] [ Fs-1 Fs]

excitation frequency range segment

Figure 5.1 Excitation frequency domain partitioning for (direct or modal) frequency
response analysis of a whole finite element model

For SOL 108, if there are many forcing frequencies, FDFREQR should be used. In
FDFREQR, the resource requirements such as disk space and memory are as large as the
resource requirements of a serial run.

47
Note that, for SOL 111, running a job with GDMODES, FDMODES, or HDMODES
automatically performs FDFREQR without requesting it explicitly. Performing
FDFREQR partitions the excitation frequencies among the number of processors defined
with the "dmp" keyword.

Example: SOL 108

For a finite element plate model that has 110 grid points and 100 CQUAD4 elements,
SOL 108 analysis is performed in parallel with the FDFREQR method. The total number
of excitation frequencies is 89 in the following entry of FREQ1.

The .f06 file describes the partitioned frequency range. In the example below, processors
1, 2, 3, and 4 execute 23, 22, 22, and 22 excitation frequency ranges. Once each
processor finishes its own analysis, the master processor collects the results and prints the
output.

48


49
5.2 DMODES + FDFREQR for SOL 111
In SOL 111, the FDMODES, GDMODES, HDMODES, RDMODES methods will
automatically continue to perform FDFREQR after the eigenvalue analysis is finished, so
you do not need to request FDFREQR explicitly for the modal frequency response.
Therefore, the parallel SOL 111 job can be run in exactly the same ways as the parallel
SOL 103. The DMP tasks in SOL 111 are described in Figure 5.2. The method selection
of DMP normal mode analysis is similar to that of SOL 103, which was introduced in
Figure 4.6.

SOL 111

FINITE ELEMENT
MODEL

DMP normal mode analysis

RDMODES / FDMODES / HDMODES /GDMODES
Depends on NREC, NUMSEG, NCLUST

FDFREQR

Figure 5.2 DMP modal frequency response analysis (SOL 111)

Example: SOL 111 (FDMODES + FDFRQR)

Each processor computes 94, 61, 26, and 27 eigenvalues, respectively, for the whole
finite element model. For the modal frequency response analysis, 19 excitation
frequencies are split to 5, 5, 5, and 4, and distributed to each processor.

50


5.3 DMODES + Serial transient calculation for SOL 112

51
In the parallel SOL 112 run, once the eigenvalue analysis is performed in parallel, the
modal transient response analysis is executed in serial. During the modal transient
response analysis, all slave processors are idle. The DMP task SOL112 is described in
Figure 5.3. Note that transient response analysis (TRD1 module) does not benefit from
DMP. The method selection of DMP normal mode analysis is similar to that of SOL 103,
which is introduced in Figure 4.6.

SOL 112

FINITE ELEMENT
MODEL

DMP normal mode analysis

RDMODES / FDMODES / HDMODES /GDMODES
Depends on NREC, NUMSEG, NCLUST

Serial
transient
calculation

Figure 5.3 DMP modal transient response analysis (SOL 112)

Example: SOL 112 (GDMODES + Serial transient calcuation)

The output of responses in the .f06 file is printed only on the master processor.

52


5.4 DMODES + serial optimization process for SOL 200

In the optimization procedures, several different types of analyses may be involved. If
any eigensolution is required, the eigenvalue analysis may be performed in parallel with
the computational methods FDMODES, GDMODES, and HDMODES used as in SOL
103.

In each iteration of the optimization procedure, the optimization algorithm runs in serial
on each processor except for the eigenvalue analysis. Whenever the eigenvalue analysis
is performed in parallel, the collected and merged eigensolutions from the master
processors are broadcast to all processors. This approach results in the same optimization
results for every processor, because all processors proceed with the optimization using
the same eigensolutions inside the optimization loop. The DMP task of optimization
procedure is described in Figure 5.4. The method selection of DMP normal mode
analysis is similar to that of SOL 103, which is introduced in Figure 4.6.

53
 SOL 200

INITIAL
DESIGN

ANALYSIS

DMP normal mode analysis

RDMODES / FDMODES / HDMODES /GDMODES

Depends on NREC, GPART, NCLUST

OPTIMIZATION

IMPROVED
DESIGN

Figure 5.4 DMP design optimization (SOL 200)

Example: SOL 200 (DMODES + Serial optimization process)

The following .f06 is a typical output format in SOL 200 with DMP, in which all
processors print the same output.

54


55


56
Chapter 6 Methods for Multi-step Nonlinear Solution

The solution sequence, SOL 401, is a multi-step solution which enables to execute a
combination of subcases listed in Table 6-1. Some of the analysis types can take
advantage of DMP solution methods. This chapter summarizes these DMP computational
methods. The supported subcases are

 ANALYSIS=STATICS static analysis (see section 6.1)

 ANALYSIS=MODES normal modes analysis (see section 6.2)
 ANALYSIS=CYCMODES cyclic modes analysis (see section 6.3)
 ANALYSIS=FOURIER Fourier modes analysis (see section 6.3)

The DMP computational methods for each subcase type are first introduced in this
chapter. Next, multi-step solution examples which include Fourier modes analysis, static
analysis, and normal modes analysis, are described in section 6.4.

Table 6-1 SOL401 Subcase Analysis Types

Analysis type Meaning
STATICS Static analysis
MODES Normal modes analysis
PRELOAD Bolt preload analysis
CYCMODES Cyclic modes analysis
FOURIER Fourier modes analysis

For more information about SOL 401, see the Simcenter nastran Multi-Step Nonlinear
User’s Guide.

6.1 Geometric Domain Static analysis (GDSTAT) in SOL 401

GDSTAT is a parallel solution for the linear static analysis of large models as introduced
in section 3.1. Simcenter nastran supports the GDSTAT method in SOL 401 for the fast
static analyses. The performance of GDSTAT depends on the size of the boundary
produced by the graph-based (GPART = 1) domain partition. When the boundary size is
small, GDSTAT is most efficient. For most cases, DMP = 2 or DMP = 4 is sufficient.

You can use system cell 649 to control GDSTAT in SOL 401.

57
Sys 649 = 0 (default) Does not select GDSTAT. Simcenter nastran performs serial
processing.
=1 Select GDSTAT with null columns in the stiffness matrix.
=2 Select GDSTAT after null columns are removed from the
stiffness matrix.

Users are recommended to use sys649 = 1 for most parallel processing cases since the sys
649=2 option spends additional time to remove null columns from the stiffness matrix.
However, for models with large open contact patches which include large numbers of
null columns, you may see a performance increase with sys649 = 2.

Example: GDSTAT for SOL 401 static subcase

For a finite element model that has 477 grid points for 4 CHEXA elements, 16 CPENTA
elements, 24 CPYRAM elements and 106 CTETRA elements, the analysis is performed
in parallel with the GDSTAT method. The following information in the .f06 describes
the detailed statistics of partitioning for the 1st subdomain (master processor):

Note that this information is printed for each time the GDSTAT method is used. The .f06
file of the slave node has the corresponding information for the 2nd subdomain:

6.2 Normal modes analysis subcase in SOL 401

Similar to a SOL 103 normal modes analysis, the GDMODES and RDMODES solution
methods can be used with the SOL 401 normal modes subcase (ANALYSIS=MODES). The

58
commands to enable GDMODES and RDMODES in SOL 401 are the same as for a SOL
103 normal modes analysis.

Example: GDMODES for SOL 401 normal modes subcase

For a finite element model that has 54 grid points and 60 solid elements with one normal
modes subcase, the analysis is performed in parallel with the GDMODES method. Once
each processor finishes its own analysis, the master processor collects the results and
prints the output in the .f06 file.

59
Example: RDMODES for SOL 401 normal modes subcase

In addition to the GDMODES method, the RDMODES method also can be applied to the
normal modes subcase. With the same example describe above, you can run RDMODES
by using keyword (auto selection option). The number of partitions is presented
in the .f06 file. Please note that RDMODES uses FDMODES for the reduced problem
which is also shown in .f06 file.




60



6.3 Cyclic modes or Fourier modes analysis subcase in SOL 401

For the cyclic modes subcase (ANALYSIS=CYCMODES) and the Fourier modes subcase
(ANALYSIS=FOURIER), a cyclic/Fourier modal solution is computed for each harmonic
index independently. Simcenter nastran takes advantage of the independent computation
as the parallelism opportunity. The modal solutions for the requested harmonic indexes
are split into processors as evenly as possible. The modes for each harmonic index are
calculated on the assigned processor. Note that each processor shares the full model
information.

Example: Cyclic modes analysis subcase

A parallel implementation of a cyclic modes subcase is performed for a finite element

model that has 173 grid points and 271 CTETRA elements. The subcase is described
below. There are a total number of 5 harmonics. The harmonic IDs are 0, 1, 2, 3, and, 4.
This model requests 10 modes to be computed on each harmonic index. Note that
harmonic IDs 0 and 4 are each a single analyses, and therefore 10 modes will be
calculated. On the other hand, the harmonic IDs 1-3 are double analyses and 20 modes
will be computed.

In the example, DMP=4 was requested. The master processor computed the cyclic modes
for harmonic IDs 0 and 1. The cyclic modes solution for harmonic IDs 2-4 was assigned
to the remaining 3 processors. Once each processor finished its own analysis, the master
processor collects the results and prints the output.

61


62


Below is the printout for the 2nd processor which had computed the modal solution for
harmonic ID 2. Other processors have similar output for their assigned harmonic indexes.

Example: Fourier modes analysis subcase

63
In this Fourier modes subcase example, the finite element model consists of 22 grid
points, 10 triangular chocking elements, and 5 quadrilateral chocking elements. The
HOUTPUT case control requests all harmonics specified in the HARMONICS case
control to output modes. Two harmonics are requested with IDs 0 and 1. In this case,
using two processors is a reasonable choice.

The .f06 file presents the “ ” and “ ” messages. Processors

1 and 2 executed serial Lanczos for harmonic IDs 0 and 1, respectively. After each
processor has finished its own analysis, the master processor collects the results and
prints the output. This model requests 2 modes to be computed on each harmonic index.
Please note that harmonic ID 0 is a single analysis, therefore 2 modes were calculated.
The harmonic ID 1 is a double analysis, therefore 4 modes were computed.

64


The slave processor computes the harmonic ID 1 and presents the eigenvalue summary
table as shown below.

6.4 DMP for a Multi-step example

65
The previous three sections discuss the case-by-case scenarios for each individual
subcase which DMP method supports. Those cover four types of subcases – the static
subcase (ANALYSIS=STATICS), the normal modes subcase (ANALYSIS=MODES), the cyclic
modes subcase (ANALYSIS=CYCMODES), and the Fourier modes subcase
(ANALYSIS=FOURIER). All other subcases listed in Table 6-1 are implemented in
sequential fashion.

The following is a multi-step example to illustrate how the DMP method can support a
combination of subcases.

Example: a Multi-step example

The model from the Fourier modes subcase in section 6.3 includes a static and normal
modes subcases in sequence. That is, DMP Fourier modes analysis is followed by a serial
static analysis (default sys649 = 0) and DMP normal modes analysis.

In the Subcase ID 1, a “ ” message is printed in the .f06 file. As

discussed in the previous section, processors 1 and 2 execute serial Lanczos for
Harmonic IDs 0 and 1, respectively. After each processor finishes its own analysis, the
master processor collects the results and prints the output.

In the Subcase ID 2, a “ ” message is printed in

the .f06 file. All processors compute the same static solution in serial fashion.

In the Subcase ID 3, a “ ” message is printed in the .f06 file. The 2-way

DMP with GDMODES is used to compute the 2 modes.

66


67


The slave processor computes the harmonic ID 1 and prints out the eigenvalue summary
table as shown below for the Subcase ID 1. The slave processor has the same nonlinear
iteration output as the master node for Subcase ID 2. Then it solves the partial solution of
the 2nd domain for GDMODES in Subcase ID 3.

68
 




69
Chapter 7 Performance Study

7.1 Performance of Parallel Processing

The performance of a parallel processing is rated by its speedup or efficiency. The
speedup Sp is defined as:
T
Sp  s  E* p
Tp
where Ts is the time to run a serial job, and Tp is the time it takes to run the same job
with p processors.
The efficiency E is defined as:
Ts
E
( p * Tp )
For better performance, it is helpful to minimize the communication overhead by using
high speed network switches. The ideal is to have high bandwidth and low latency.

7.2 Industrial Case Study 1

A trimmed car body FE model with SOL 103 is used as a case study to analyze the
technologies presented in Simcenter nastran. This type of car model has all major
components of the car, such as wheels, engine, etc. incorporated. Tables 6-1 and 6-2
present the details for this finite element model.

Table 7-1 Model statistics of trimmed car body FE model

Number of Number of Number of Number of

Nodes shell elements solid elements rigid elements
380,007 361,249 3,762 9,056

Table 7-2 The size of sets for the trimmed car body FE model

The size of set

g-set n-set f-set a-set

2,280,042 2,223,139 2,223,109 1,937,282

70
Table 6-3 demonstrates the effect of the automated geometric domain decomposition on
this model by showing the number of interior and boundary nodes of the partitions.
Several observations can be made. The interior range size depends on the quality of the
automated partitioning. The boundary size increases with the number of partitions.
Finally, the boundary size is at least two orders of magnitude smaller than the interior
size, which is important to the computational efficiency.

Table 7-3 Results of geometry domain decomposition

Number of Maximum Minimum Maximum Minimum

Partitions Interior Interior Boundary Boundary

2 197,460 182,199 354 354

4 104,788 90,791 572 418
8 54,496 40,067 518 450

The automated frequency domain decomposition results are shown in Table 6-4. There
are 840 modes in the frequency range of interest.

Table 7-4. Frequency domain decomposition statistics

Number of Minimum number of Maximum number of

segments modes among segments modes among segments
2 380 460
4 193 252
6 133 165
7 105 141
8 92 119

The automated geometric domain partitioning techniques usually provide only even, and
preferably binary, numbered domains. This is because these techniques are primarily
based on binary graph partitioning. This is not a restriction, as shared memory
workstations tend to have an even number of processors. On the other hand, in the
frequency domain decomposition, odd numbers of segments are also allowed. This
technology is insensitive to that issue and enables the use of odd-numbered workstations
via the hierarchic technology in a workstation cluster environment.

71
The analysis was executed on a cluster of eight workstations, each containing eight
processors with a 1.5 GHz clock cycle. The cluster had a one gigabyte Ethernet network
connection. The option gpart=1 is used.

Table 7-5. Execution times on workstation cluster with HDMODES

Number of Number of Number of Elapsed

I/O (GB) Speedup
processors partitions segments time[min:sec]
1 1 1 1,028.4 528:58 1.00
4 2 2 431.4 167:15 3.13
8 4 2 266.9 83:41 6.26
16 8 2 191.1 45:16 11.57
32 8 4 98.3 34:07 15.35
48 8 6 77.8 27:14 19.23
56 8 7 67.1 24:41 21.22
64 8 8 61.4 27:00 19.40

The task of finding the natural frequencies and mode shapes of such a model is an
enormous one. It is an overnight job with more than a terabyte of I/O operations. The
execution on a single processor is impractical considering the work environment and time
schedule at automobile companies.

Fig. 7.1 and Fig. 7.2 show the elapsed time and disk I/O of HDMODES with 8 geometry
partitions for different numbers of frequency segments. With 56 processors, 8 geometry
partitions in each workstation and 7 frequency segments across workstations are used.
The elapsed time for 32 processors is already a practical execution. The efficiency above
decreases, but the speedup is still increasing. It peaks at 56 processors, although a wider
frequency range for this model may extend that peak to 64 or higher. It means that 7
frequency segments with 8 geometry partitioning for each frequency segment is the most
suitable partitioning for this test FE model in this workstation cluster environment.

72
 HDMODES
(8 geometry partitions)

600

550

500

450

400
Elapsed time (min.)

350

300

250

200

150

100

50

0
1 16 32 48 56 64
Number of processors

Fig. 7.1 Elapsed time of HDMODES with 8 geometry partitions for different numbers of
frequency segment.

73
 HDMODES
(8 geometry partitions)

1100

1000

900

800

700
Disk I/O (GB)

600

500

400

300

200

100

0
1 16 32 48 56 64
Number of processors

Fig. 7.2 Disk I/O of HDMODES with 8 geometry partitions for different numbers of frequency
segment.

7.3 Industrial Case Study 2

A trimmed car body FE model with SOL 103 is used as a case study to analyze the
performance of RDMODES in Simcenter Nastran. This model has g-size around 20
million and f-size about 10 million. The number of grid points is around 3.6 million, and
the number of ctetra elements is about 2.3 million.

The example was run through RDMODES with rdsparse (sparse eigenvector recovery)
turned on and nrec=256.

Scalability – elapsed time vs. number of processors

It computed modes up to 10,000 Hz. The following table and graph will show the
summary of the performance from one processor to 64 processors.

74
 Table 7-6. Execution times on workstation cluster with RDMODES

Number of Elapsed time I/O CPU speedup

processors (min:sec) (G-bytes) (seconds)
1 4370:52 9110.7G 186277 1.00
2 2548:23 5136.2G 110519.1 1.78
4 1402:05 2711.5G 58795.8 3.12
8 1019:13 1431.5G 29086.3 4.29
16 678:09 945.6G 19272.7 6.45
32 505:36 637.8G 10731.6 8.64
64 354:53 526.3G 8920.1 12.32

RDMODES for 20M G-size Car Model up to 10,000 Hz (about 260 modes)

300,000

250,000
elapsed time (sec)

200,000

150,000 rdmoes

100,000

50,000

0
1 2 4 8 16 32 64
number of processors

Fig. 7.3 Elapsed time of RDMODES with 256 geometry partitions for the different
number of processors.

The analysis was also executed on a Linux cluster of 64 nodes; each processor is 1.8 GHz
clock cycle. The cluster had a one gigabyte Ethernet network connection.
Due to the large dimension of this model, the task of finding the natural frequencies and
mode shapes of such a model is more enormous. The elapsed time is saved significantly
through RDMODES with more processors.

Table 7-6 and Figure 7.3 showed that, when 64 processors are used, the computation is
speedup by 12. The elapsed time is around 6 hours with 64 processors, while the time of
one processor is 73 hours.

75
Scalability – elapsed time vs. number of modes
The following table and graph show multi-level RDMODES runs (dmp=64, nrec=256)
with respect to frequency range up to 10000, 20000, 30000, 40000 and 50000 Hz. It took
about 110 extra minutes and found about 2900 more modes by increasing frequency
range from 10000 to 50000. It clearly demonstrates that the RDMODES is capable for
computing large number of modes.

RDMODES
Modes
below (Hz) Elapsed Number
time of modes
10000 354:53 295
20000 430:47 764
30000 441:20 1453
40000 455:09 2276
50000 462:04 3255

Mode scaling

30000

25000

20000
elapsed time

15000

10000
rdmodes
5000

0
10000 20000 30000 40000 50000
find modes up to frequency (Hz)

76
Chapter 8 Installation and Configuration of DMP

8.1 Overview

Simcenter nastran offers the ability to run certain solution sequences in parallel using the
Message Passing Interface (MPI), an industry-wide standard library for C and Fortran
message-passing programs. MPI programs can be run on SMP computers, NUMA
computers, distributed computers, and any collection of computers supported by the MPI
package.

Note: Further information on MPI can be obtain online at

http://www.mpi-forum.org

MPI is included with the installation of Simcenter nastran.

Special Considerations

To install Simcenter nastran for Distributed Memory Parallel (DMP) operations, you
must select one of the following three installation schemes if you want to use more than
one host in a single Simcenter nastran job:

 Install Simcenter nastran on a filesystem that is global to every host. This

provides the easiest installation and system administration, but may present
network load issues when the Simcenter nastran is started and the delivery
databases are being read.
 Install Simcenter nastran on every host on host-private filesystems. This is harder
to install and administer, but reduces the network load when Simcenter nastran is
started.
 A combination of the above.

Note: In all cases, the nastran command must have the same pathname, or be in the
default PATH of every host that will run a DMP job. Recall that your “.profile” and
“.login” files are not used for rcp(1) and rsh(1) operations.

8.2 Requirements

Platform MPI
X86_64 Linux Intel MPI is included with the Simcenter nastran installation, and is
automatically invoked when a DMP job is executed on Linux. Intel
MPI has specific requirements, for example, Python must be
installed. These requirements are documented in the Intel MPI
release notes found at:

77
 https://software.intel.com/en-us/articles/intel-mpi-library-
documentation
Windows-64 Intel MPI is included with the Simcenter nastran installation. See the
Windows Single Host Instructions.

In the descriptions that follow, the “local” node is the computer you issue the nastran
command on, the “master” node is the first computer named by the “hosts” keyword, and
the “slave” nodes are the remaining systems listed in the “hosts” list.

The following are some general requirements for running Simcenter nastran DMP jobs:
 Simcenter nastran must be properly installed on all the hosts listed by the “hosts”
keyword.
 On Linux, either rsh or ssh can be used as a remote command. Secure Shell (ssh)
is supported on Linux provided that:
a. The environment variable MPI_REMSH is set to ssh.
b. The argument s.rsh=ssh is included on the nastran command line.
c. The argument s.rcp=scp is included on the nastran command line.
 You must have access to each system you want to access in a distributed job. For
example, when using rsh, you can test this with the command:
rsh <node> [-1 <username>] date
where <node> is the name of the node and <username> is an alternate username
on the remote system if your current username is not valid. For example:
rsh node1 date
The output from the above command should be in a single line containing the
current date on node1 in a format similar to
Thu Jul 17 13:06:49 EST 2003
If any other output is present, you should determine the source of the output and
correct the problem. If you cannot eliminate the output, you will not be able to use
the distributed execution capabilities of the nastran command.
 You must have “remote execution” privileges on all the hosts listed by the “hosts”
keyword. That is, a password must not be required to execute a remote copy (rcp)
or remote shell (rsh or remsh) command. See your system administrator for
information on this.
 The input data file must be accessible on the local host.
 INCLUDE files must be local-to, or visible-from, each host.
 All default output files, i.e., those without ASSIGN statements, will be written to
a directory accessible to the local host.
 The scratch directory can be a global or local file system. Your scratch directory
should be local to each host, i.e., you specify per-host “sdirectory” values.
 The pathname of the nastran command must be the same on all hosts, or on the
default PATH of each host, used in the analysis.
 If you execute a restart, you must specify the identical values for “dmparallel” and
“hosts” as were used on the cold start.
 In a restart, i.e., a job that uses an existing database, the DBSets must be local-to,
or visible-from, the remote system.

78
Note: Recall that remote executions do not run a “login” shell. That is, your “.profile” or
“.login” script is not executed.

When running a DMP job, nastran keywords are processed on both the local and
master/slave systems. Keywords that control the job’s output and interaction with you are
processed on the local system. These are:
Keyword Description
append Requests the .f06, .f04, and .log files to be concatenated.
dmparallel (or dmp) Specifies the number of tasks for a Distributed Memory Parallel
(DMP) analysis. This value may only be set on the command line.
gpart Selects the geometry partitioning option for a hierarchic dmp
(HDMP) solution.
hostovercommit Requests more tasks per host than CPUs.
hosts Specifies list of hosts to use. Separate hosts with the PATH
separator, i.e, ;.
mergeresults Specifies the results from each DMP task are to be merged into
the standard files from the master host.
nclust Specifies the number of frequency segments for a hierarchic dmp
(HDMP) solution.
ncmd Specifies an alternate notification command
notify Requests notification when the job completes.
old Specifies versioning or deletion of previously existing output
files.
oldtypes Specifies additional user file types to be versioned or deleted.
out Specifies an alternate output file prefix.
rcmd Specifies the nastran command path on the master/slave systems.
scratch Specifies the database DBSets are to be deleted at job completion.
sdirectory Specifies each per-host directory to contain Simcenter nastran
temporary files. Separate directories with the PATH separator.
slaveout Specifies the .f04 and .f06 files from the slave tasks are to be
appended to the .f04 and .f06 files of the master task.
xmonitor Requests XMONITOR to monitor the master task’s progress.

The “sdirectory” keyword is special, as the command line, RC files on the current host,
and RC files on the each master and slave host will all be considered when establishing a
scratch directory. All remaining keywords are only scanned on the master and slave
systems.

Once “dmparallel=number” is processed, the following processing takes place:

1. Process the RC files on the local system if the “version” keyword has been
defined in the command initialization file or the command line.
2. Process the RC file specified by the “rcf” keyword if it was defined on the
command line.
3. Determine the full pathname of the input file so that its visibility from the master
and each slave host can be tested.

79
4. Create a “touch” file in the specified output file so that its visibility from the
master and each slave host can be tested.
5. If the “dmpdeny” utility, i.e., install_dir/nxnr/arch/dmpdeny, exists and is
executable, run it, and save its output.
6. If the “dmpaccept” utility, i.e., install_dir/nxnr/arch/ dmpaccept, exists and is
executable, run it, and save its output.
7. Ensure “scratch=no” was set if the “dbs” keyword was set.
8. Determine every possible pairing of host and sdirectory by scanning each list in a
round-robin order. That is, the first host is paired with the first sdirectory, the
second host with the second sdirectory, and so on.
9. Execute the following steps for each host-sdirectory pair determined above until
host-sdirectory pairs have been assigned to each of the tasks requested by the
“dmparallel” keyword or no more host-sdirectory pairs are available. Steps 9a.
through 9f. are executed only once per host-sdirectory pair.
a) Verify that host exists and you are able to run a command on that
system.
b) If the “rcmd” keyword was specified, attempt to execute that
command on host, display an error and cancel the job if it fails.
Otherwise attempt to execute the pathname of the current nastran
command on host. If it fails, attempt to execute the basename of
the current nastran command on host. Display an error and cancel
the job if both checks fail.
c) Run the remote nastran command identified in the previous step to
determine: if the input data file is visible; if the “touch” file is
visible, if the “sdirectory” (if identified on the local system) exists;
if the “dbs” directory (if identified on the local system) exists; the
“sdirectory” value in the RC files defined on host; and finally the
numeric format of host.
d) Drop this host-sdirectory pair from further consideration if a
scratch directory was identified on the command line or in a local
RC file, but does not exist on host.
e) Display an error and cancel the job if the numeric format of host
differs from the numeric format of the local host.
f) Display an error and cancel the job if the directory specified by a
“dbs” keyword on the command line or in a local RC file does not
exist on host.
g) Assign the current host-sdirectory pair to the next task; save the
per-host visibility flags, “rcmd”, and “sdirectory” values.
10. Display an error and cancel the job if one or more of the tasks requested by the
“dmparallel” keyword have not been assigned.
11. Delete the “touch” file created above.
12. The remaining steps are done in a background process (possibly some time later)
if “batch=yes” or “after” was specified.
a) Copy the input data file to the scratch directory of any host that
could not see the input data file.

80
 b) Set “out” to the host-specific scratch directory value of every host
that could not see the output directory.
c) Copy the remaining keywords on the command line that were not
processed, to a local RC file in the scratch directory on the remote
node.
d) Run the DMP job using the system’s MPI startup command. Note
that each task will write its files to task-specific names.
e) Process the “old” and “oldtypes” keywords on the local node.
f) Copy the output files (.f04, .f06, .log, .ndb, .pch, .plt) from the
master task to the directory specified by the “output” keyword and
delete the files from the master node if it could not see the output
directory.
g) Process the “append” keyword on the local node.
h) Process the “notify” keyword on the local node.

Once the job has completed, the .f06, .f04, .log, .ndb, .op2, .plt, .pch, and .xdb files from
the master task will be present as if the job were run locally.

Note: No attempt is made to copy DBSet files between the local and master/slave
systems. If this is required, you must handle this yourself and set the “dbs” keyword
appropriately.

8.3 Windows Single Host Instructions

The following instructions describe how to install Windows DMP on a single host which
includes multiple cores.

Step 1: Before you can follow the remaining steps, an administrator will need to perform
the following tasks.
 You must have a login directory that you own.
 You must have full permissions to all working directories and scratch directories.
 Your working directory must be shared, for example, D:\workdir shared as
\\host\workdir.

Step 2: Install Simcenter nastran to the Windows host

 Install Simcenter nastran to a directory with no spaces in the path. For example,
do not install to the default path of C:\Program Files. Spaces in the path will
prevent environment variables from being properly interpreted.

Step 3: Set up the Intel MPI Windows Service

 Login as the local administrator and open a DOS shell.
 Go to the bin directory:
cd /d …installation_path\nxnr\em64tnt\impi\bin
 Enter the following:
hydra_service.exe –install

81
Step 4: Unset PLATFORM_MPI
 The environment variable PLATFORM_MPI was required in previous releases
to run DMP. It must be unset if it is still defined on your system. Enter the
following to check if it is defined:
echo %PLATFORM_MPI%
%PLATFORM_MPI% will return if it is undefined.
If it is defined, a file system path will appear.
Enter the following if you need to unset the variable:
set PLATFORM_MPI=

Step 5: Each user must set up password-less MPI in a DOS shell

 Go to the bin directory:
cd /d …installation_path\nxnr\em64tnt\impi\bin
 Enter the following:
mpiexec.hydra.exe –register
 Follow the prompts to cache your password.
 Repeat these steps whenever you change your password.

Step 6: Define MPI_ROOT and update PATH

 Define the MPI_ROOT environment variable:
set MPI_ROOT=installation_path\nxnr\em64tnt\impi
 Include the following ‘bin’ directory in your PATH variable:
set PATH=installation_path\nxnr\em64tnt\impi\bin;%PATH%

Step 7: Test Windows DMP with a .BAT script

 Use a text editor to create the following .BAT file, replacing hostname with the
actual host name. The example assumes a ‘workdir’ and enabled sharing for this
folder, and a Simcenter nastran installation location of D:\NXNr. Note the use of
quoted, forward slashes for the shared directory.
Name the file dmptest.BAT:
@ECHO OFF
set NXN_BASE=D:\NXNr
%NXN_BASE%\bin\nastran.exe ^
“// hostname /workdir/plan10g.dat” ^
out=”// hostname /workdir” ^
mem=130mw dmp=2 hosts= hostname ^
sdir=D:\Scratch slaveout=yes scr=yes
 Open a DOS shell, and change directories to ‘workdir’:
cd \\hostname\workdir
 Copy the file plan10g.dat from the tpl directory:
installation_path\nxnr\nast\tpl\plan10g.dat
to //hostname/workdir/
 Execute the .BAT file:
dmptest.bat
 The results should look similar to the following:

82
8.4 Windows Multiple Host Instructions (True Cluster)
The following instructions describe how to install Windows DMP across multiple hosts
(in a cluster). See the Windows Single Host Instructions to install and run Windows DMP
on a single host which includes multiple cores.

General Requirements
 Simcenter nastran must be installed in the same location on each node and must
be accessible to each user.
 Each user must have full permissions to all working directories and scratch
directories.
 All working directories must be shared to all nodes.
 Only Server 2008 and Server 2012 are supported in the true cluster mode.
 DMP Jobs must be launched from a DOS shell.

Step 1: SUA Removal Instructions

Since June 1, 2013, Microsoft has removed support for SUA and recommends that
Cygwin be installed in its place:
http://technet.microsoft.com/library/hh831568.aspx
If SUA is installed on your cluster, it must be removed before installing Cygwin 64-bit.
You can follow the following SUA removal instructions. Note that the SUA removal
instructions must be repeated on each node of the cluster.
If you do not have SUA installed, skip forward to the Cygwin 64-bit installation
instructions in Step 2.

A. Stop the SSHD service

 Login as Administrator

83
  Start Computer Management (Start, Search, “Computer Management”)

 Select Services, right-click the sshd service, and select stop

 When the service is stopped, close the Manage window

B. Remove the SSHD service

 Login as Administrator
 Open a Command Prompt and cd to C:\Windows\System32
 Remove the SSHD service with the command
sc delete “sshd”

C. Remove the SUA SDK

 Login as Administrator
 Open Control Panel, select Uninstall a program
 Select “Utilities and SDK for UNIX-based Applications”, and select Uninstall

84
D. Remove SUA as a Windows Feature
 Login as Administrator
 Open Control Panel, select Programs, then Programs and Features
 Select Turn Windows features on or off
 Un-check Subsystem for UNIX-based applications, and select OK

 When asked to reboot the computer, select Restart Now.

E. Verify that SUA is no longer a part of any environment variable.

 After the previous re-boot, login as Administrator.
 Right click on the desktop Computer icon, select Properties, Advanced System
Settings.
 Select Environment Variables and edit out any references to SUA for both User
and System.

F. Remove the SUA directory (optional)

 After the previous re-boot, login as Administrator

85
  Right click on the SUA directory, and select Delete

Step 2: Cygwin 64-bit Installation

Prerequisites
 Cygwin 64-bit is required for Simcenter nastran
 If a previous version of Cygwin is installed, it should be removed before installing
Cygwin 64-bit.
 Instructions for removing previous versions of Cygwin can be found at
http://cygwin.wikia.com/wiki/Uninstalling_Cygwin
 Simcenter nastran requires Cygwin-64 to be installed to the C:\ drive under
C:\cygwin64.
 Cygwin-64 must be installed on each node of the cluster

Begin the Installation

A. Login as local Administrator

B. Go to http://cygwin.com/install.html, right-click “setup-x86_64.exe”, choose save as

target. Save to a local directory, such as C:\Temp.

C. Double-click on C:\Temp\setup-x86_64.exe to begin the installation. Select Next

on the setup window:

D. Select “Install from Internet”, then Next.

86
E. Make sure the installation directory is C:\cygwin64, select “All Users”, then Next.

F. Select a suitable scratch directory to hold installation files, such as C:\Temp. Select
Next.

G. Select your internet connection, then Next.

H. Choose a download site, for example http://mirrors.kernel.org. Select Next.

I. From Select Packages:

 Expand Shells, select tcsh and mksh.

87
  Expand Archive and select unzip and zip.

 Expand Net category and select openssh and openssl.

Additional packages can be selected later after the installation. Select Next to
continue.

J. A window will open listing dependencies. Select Next to resolve dependencies.

K. A window will open to show the installation progress. Select Finish after the
installation completes.

Step 3: Cygwin 64-bit Configuration

88
A. Right-click the Computer icon, select Advanced Settings, then Environment
Variables. Add C:\cygwin64\bin to the global PATH in the System
Variables.

B. Administrator setup of SSH service

Login as Administrator, open a Cygwin64 terminal, and issue the command:
ssh-host-config
You will be prompted to answer several questions:
Questions Answers
Should privilege separation be used? Yes
Should this script create a new local account ‘sshd’? Yes
Do you want to install sshd as a service? Yes
Enter the value of CYGWIN for the daemon. ntsec
Do you want to use a different name for the SSH service? This is optional, but
it is simplest to enter
No.

Once the setup of the SSH service is complete, you should see something similar
to the following:

Note: If you plan to use a Management Agent, then follow the instructions given
here:
http://docs.oracle.com/cd/E24628_01/install.121/e22624/preinstall_req_cygwin_s
sh.htm

C. Startup the SSH service from a Cygwin64 terminal with:

cygrunsrv –S sshd

Once the service starts, the command prompt should return without error. The
service will also automatically start after each reboot. You can also verify that the

89
 service is running by opening Control Panel, System and Security, and Selecting
Administrative tools, and clicking on Services.

D. Logout from the local Administrator account, then login as one of the cluster
users. Double-click on the Cygwin64 Terminal icon from the desktop. Note
that a new home directory is created as each user logs in.

Verify that your shell shows correctly with

‘echo $SHELL’ (as shown below).
Also verify that ssh is working, and that you can run ssh on the localhost (as
shown below).

Note: The Administrator can edit /etc/passwd to change the shell assigned to each
user, after each user has logged in and their /home directory has been created.

E. Final Cygwin 64-bit Configuration Steps

 For each user of the cluster, repeat the previous step until all users have a
home directory on the local node, and each can run ssh on the local node.
(Instructions to setup password-less ssh are provided later.)
 If necessary, the local Administrator can open a Cygwin64 terminal and
edit /etc/passwd to define the appropriate shell for each user.
 Each user’s login directory must exist on all nodes with normal
permissions to the user
 Each user’s shell should be set (for example, to /bin/ksh)
 Repeat all of the above installation instructions for Cygwin 64-bit on each
additional node in the cluster.

After the installation and configuration of Cygwin 64-bit on each node of the cluster, the
final step below shows how to setup password-less ssh for each user, install the latest
version of Simcenter nastran, and configure MPI on the cluster for each user.

Step 4: Remaining Installation Tasks for Simcenter nastran DMP

90
The remaining tasks to be completed in Step 4 are:
A. Set up password-less SSH with Cygwin 64-bit for each user
B. Install the latest release of Simcenter nastran or modify an existing release
C. Set up Intel MPI Windows services
D. Set up password-less MPI
E. Test Windows DMP with a .BAT file from a DOS shell

Note: If a previous version of Nastran with HP MPI is installed on the cluster, there is no
need to remove it. It is not used with Intel MPI and Simcenter Nastran.

A. Set up password-less SSH with Cygwin 64-bit for each user

Each user must open a Cygwin64 terminal, which will place them in their home directory.
 Enter the command:
ssh-keygen –t dsa (or, alternately with –t rsa)
 Press Enter when you are presented a prompt.
A text image will be displayed when ssh-keygen is done:

 Enter the .ssh directory with: cd .ssh

 Display .ssh directory contents with: ls –alt
 Copy the public key with: cp id_dsa.pub authorized_keys
 Change permissions with: chmod 644 authorized_keys

 The first time you run ‘ssh hostname date’, answer ‘yes’ if prompted. All subsequent
ssh commands to hostname will be password-less, as shown below.

91
 The final step in this process is to provide password-less ssh between all nodes for
each user. In order to provide password-less ssh between any two nodes, each user
needs to collect all of the public keys created in the /home/username/.ssh
directories and insert them into /home/username/.ssh/authorized_keys on all the
nodes.

For example, if user harry created the following public keys on these nodes:
Master: C:\cygwin64\home\harry\.ssh\id_dsa.pub
Slave1: C:\cygwin64\home\harry\.ssh\id_dsa.pub
Slave2: C:\cygwin64\home\harry\.ssh\id_dsa.pub

Then, user harry must insert the contents of those 3 public keys into authorized_keys
on the following nodes:
Master: C:\cygwin64\home\harry\.ssh\authorized_keys
Slave1: C:\cygwin64\home\harry\.ssh\authorized_keys
Slave2: C:\cygwin64\home\harry\.ssh\authorized_keys

When all public keys for all the nodes are contained in the authorized_keys file on
each node, then any node may run ssh between any other node without being
prompted for a password.

Note that the permissions for authorized_keys must be set to 0600 from a Cygwin64
Terminal.

Also note that if a user’s home directory is shared to all of the nodes, then it is only
necessary to run ssh-keygen once, and create a single authorized_keys file in the
user’s $HOME/.ssh/ directory.
Even though a node-name will appear at the end of the security string in the public
key file (e.g., id_dsa.pub), which will be the name of the node on which ‘ssh-keygen’
was launched, it is merely a comment and is ignored by SSH.
If the home directories are not shared among the nodes, then it is necessary to run
ssh-keygen on each node and copy all of the public keys generated into the
authorized_keys file on each node.

B. Install Simcenter Nastran to each Windows node

 Install Simcenter Nastran on all the nodes of the cluster in the same location on each
node.

92
 Using your Webkey account, download the Simcenter nastran installation (.zip file)
from https://download.industrysoftware.automation.siemens.com/.

 Unzip the installation file and launch autorun. Install to any directory that does not
have spaces in the path name, for example, D:\NXNr. Do not install to the default path
of C:\Program Files(x86). Spaces in the path will prevent several Environment
variables from being properly interpreted.

C. Set up Intel MPI Windows Service on each node

 Login as local Administrator and open a Command Prompt.

 Change directory with:

cd /d D:\NXN10\nxn10\em64tnt\impi\bin

 Type the following command:

hydra_service.exe –install

 Verify that the service is running under Administrative Tools/Services

D. Set up password-less MPI for each user on each node

Each user must enter the following commands at a Command Prompt on each node of
the cluster:

set MPI_ROOT=D:\NXN10\nxn10\em64tnt\impi

set PATH=D:\NXN10\nxn10\em64tnt\impi\bin;%PATH%

cd /d D:\NXN10\nxn10\em64tnt\mpi\bin

mpiexec.hydra.exe –register

E. Test Windows DMP with a .BAT script

 Use a text editor to create the following .BAT file, replacing nodei with the actual
host names available on your cluster. Name the file dmptest.BAT.

@ECHO OFF
set NXN_BASE=D:\NXN10
%NXN_BASE%\bin\nastran.exe
"//node0/Workdir/UserName/plan10g.dat"
out="//node0/Workdir/UserName"
mem=130mw dmp=2 hosts=node0;node1
sdir=D:\Scratch slaveout=yes scr=yes

93
 Open a DOS shell, then cd \\node0\Workdir\UserName, where node0, Workdir
and UserName are replaced with correct values for your hosts.

 Copy a test file from the Simcenter nastran tpl to your current directory

copy D:\NXN10\nxn10\nast\tpl\plan10g.dat

 Execute the .BAT file:

dmptest.bat

8.5 Determining Hosts

The nastran command uses the following hierarchy to determine the list of hosts to use:

 The nastran command “hosts” keyword on the command line

 The nastran command “hosts” keyword in an RC file.
 The local host.

Consider the following example:

The following job will run on the local host:
nxnr nastran example dmparallel=4

The following job will run on the first four available nodes from the set “node1”, “node2”,
“node3”, “node4”, “node5”.
nxnr nastran example dmparallel=4 hosts=node1:node2:node3:node4:node5

The following job reads the file “my.host.list”.

nxnr nastran example dmparallel=4 hosts=my.host.list

The nastran command provides a simple host allocation method. If a host listed by the
“hosts” keyword is unavailable, it will be skipped and the next host considered. As long
as at least the number of processors specified by the “dmparallel” keyword are available
on one or more of the listed hosts, the job will be allowed to run.

Hosts on Linux
On Linux systems, the “hosts” keyword needs to specify the host name of the compute
nodes. A Linux example of job submittal is:
nxnr nastran example dmparallel=4 hosts=n1:n2:n3:n4

Using the “hosts” Keyword (Distributed Jobs Under LSF)

The “hosts” keyword will default to the value set by LSF when running as a distributed
job and no other value for “hosts” was set on the command line or in an RC file.
Example:

94
bsub -n 4 nxnr nastran example dmp=4

This job will use four hosts selected by LSF. Note, the number of tasks appears twice:
once for use by LSF, and once for use by Simcenter nastran.

Using PBS with Simcenter nastran

Portable Batch System (PBS) is a queuing system that can be used to submit Simcenter
nastran serial and DMP jobs. Once you have downloaded and installed PBS, you can use
the following sample script to run a Simcenter nastran DMP job under PBS.

The “hosts” keyword will default to the value set by PBS_NODEFILE when running as a
distributed job and no other value for “hosts” was set on the command line or in an RC
file.

#!/bin/ksh
#
# pbs_nast: PBS script to use with Simcenter nastran
#
# Usage: qsub -lnodes=Number-Of-Nodes pbs-nast
#
# Assume the data file is located in the directory whence the qsub
# command was issued.
#
dat=$PBS_O_WORKDIR/d10101d.dat
#
jobdat=${dat##*/}
#
# Change the working directory to the scratch directory.
#
TMPDIR=/scratch
cd $TMPDIR
#
# Pull the bulk data file over.
#
rcp $PBS_O_HOST:$dat .
#
# Determine the number of ranks.
#
dmparallel=$(sed -n -e ‘$=’ $PBS_NODEFILE)
#
# Build the hosts keyword value.
#
hostskwd=’sed -e :a -e ‘/$/N; s/\n/:/; ta’ $PBS_NODEFILE’
#
# Run the Simcenter nastran job. If using version 4.1 or above, comment
the following
# two lines and uncomment the next two commented lines.
#
nxnr nastran $jobdat dmparallel=$dmparallel hosts=$hostskwd \
scratch=yes batch=no
#
# If using version 4.1 or above, uncomment the following two commented
lines

95
# comment out the two lines preceeding the comment lines here
#
# nxnr nastran $jobdat dmparallel=$dmparallel hosts=$PBS_NODEFILE \
# scratch=yes batch=no
#
#
#
# Push the files back to the submitting host.
#
jobout=${jobdat%.*}
out=${dat%/*}
rcp -p $jobout.log $PBS_O_HOST:$out
rcp -p $jobout.f04 $PBS_O_HOST:$out
rcp -p $jobout.f06 $PBS_O_HOST:$out
rcp -p $jobout.op2 $PBS_O_HOST:$out
#
# END

Note: Be aware that in order to receive your job’s stdout and stderr, your .rhosts file on
the node issuing the “qsub” command must permit access from the remote host(s).

8.6 Managing Host-Database Directory Assignments

The performance of the disk subsystem containing the permanent end SCRATCH
DBSets can have a significant impact on Simcenter nastran performance. In the case of a
DMP job, the impact can be even greater if multiple tasks are using the same file system.
To allow unique directories to be assigned to each task, the “dbs”, “hosts”, and
“sdirectory” keywords are treated as lists scanned in a round-robin order. With this
feature, you can finely control the use of disk I/O access paths by your job.

The following examples show the effect of the round-robin ordering.

nxnr nastran example dmparallel=4 hosts=a:b
sdirectory=/aa:/ba:/ab:/bb dbs=/aa:/ba:/ab:/bb

This example will assign the following host-sdirectory pairs (assuming hosts “a” and “b”
each have at least two processors):
Task Host Scratch Directory DBS Directory
1 a /aa /aa
2 b /ba /ba
3 a /ab /ab
4 b /bb /bb

If directory “/ba” was not available for writing by you on host “b”, the tasks assignments
would be (assuming host “a” has at least three processors):
Task Host Scratch Directory DBS Directory
1 a /aa /aa
2 a /ab /ab
3 b /bb /bb
4 a /aa /aa

96
8.7 Managing Files
When a Simcenter nastran DMP job is running, the input file is directly read by each MPI
task that can read the file, for example, via NFS. Each host that cannot read the input file
will read a local copy of the file that is copied, via rcp(1), to the job’s scratch directory
(“sdirectory” keyword) before the job begins.

A similar check is made for the output directory. Any host that can write to the output
directory (“out” keyword) will directly write its .f04, .f06, .log and other default output
files to that directory. Any host that cannot see the output directory will write its default
output files to the job’s scratch directory. These files will then be copied, again via rcp(1),
back to the output directory at the end of the job.

Note: The nastran command will perform these tests by converting your pathname value
to an absolute pathname. As a result, a path that varies depending upon the host will be
labeled as unreadable.

If the “sdirectory” keyword is not specified on the command line or in an RC file on the
local host, each master or slave host will use its own scratch directory. This directory is
determined on the master and each slave host by examining its command initialization
file and version-specific RC files if the “version” keyword was defined.
Do not use an ASSIGN statement for any file that will be written by Simcenter nastran in
a Distributed Memory Parallel (DMP) job. Instead, use the “sdirectory” and “dbs”
keywords to specify names of the SCRATCH and permanent DB Sets.

8.8 Performance Issues

In addition to the normal performance issues associated with a serial or SMP job, a DMP
job adds communication bandwidth as a critical performance characteristic. The basic
communications channels are:
 Shared memory - SMP and NUMA systems.
 Interconnect, adapter, or switch - NUMA and distributed systems.
 High-speed special-purpose network, for example, HIPPI - all systems.
 TCP/IP network - all systems.
 Infiniband, Myrinet and Quadrics network connections on Linux are supported
out of the box.

The performance of any Simcenter nastran job depends upon CPU, memory subsystem,
and I/O subsystem performance. A Distributed Memory Parallel (DMP) job on an SMP
or NUMA system is extremely sensitive to I/O subsystem performance since each task
independently accesses the I/O subsystem.
If you select independent disks (not in RAID configuration) for your scratch drives, you
are encouraged on SMP and NUMA systems to partition your scratch directory and

97
database assignments on DMP jobs using the “sdirectory” and “dbs” nastran command
keywords.

Example 1:

nxnr nastran example dmp=4

sdir=/scr1:/scr2:/scr3:/scr4\ dbs=/dbs1:/dbs2:/dbs3:/dbs4

The following assignments will be made in this job:

Task sdirectory dbs
1 /scr1 /dbs1
2 /scr2 /dbs2
3 /scr3 /dbs3
4 /scr4 /dbs4
The preceding example will perform substantially better than the following job, which
uses the default assignments for the “sdirectory” and the “dbs” keywords.

Example 2:

install-dir\nxnr nastran example dmp=4

While the ultimate effect of the communications channel on job performance is

dependent upon the solution sequence, for best overall job performance, you should try to
use the fastest communications channels available.

8.9 Overview of Running a DMP job

This section gives a brief overview of running DMP jobs. See the other sections of this
guide for more complete details on this topic.

Command-line Syntax

You can start a Simcenter nastran DMP job using the command:
$ nxnr nastran example dmp=2 hosts=n1:n2
where “n1:n2” indicates the hosts to be used in the run.
Valid output is:
Determining available hosts, please wait...
DMP task 1: host="ugsclust1" sdir="/scratch" dbs="/scratch/plan10g";
DMP task 2: host="node1.local"; sdir="/scratch";
dbs="/scratch/plan10g";
Simcenter nastran beginning distributed job plan10g.
Simcenter nastran V4.0 (Intel Linux 2.4.20-20.7smp) Mon May 23 11:21:37
2005
Simcenter nastran V4.0 (Intel Linux 2.4.20-20.7bigmem) Mon May 23
11:21:37 2005
Simcenter nastran beginning child job plan10g.T22389_37.t1 on
node1.local.
Simcenter nastran beginning child job plan10g.t0 on ugsclust1 (master).

98
Note: The “beginning child job” lines may appear in a random order.

Error Examples
The following error is an authorization problem.

*** USER FATAL MESSAGE 3060 (PREFACE)

SUBROUTINE MODEL - OPTION NAST NOT IN APPROVED LIST.
SYSTEM DATE (MM/DD/YY): mm/dd/yy
SYSTEM UGSID: n (DECIMAL) n (HEXADECIMAL)

Likely causes are:

 The license or authorization file does not include the ability to make DMP runs.
 The license or authorization file was not accessible to the first node in the hosts
list.

99
Chapter 9 GPGPU computing

9.1 Overview
The graphics processing unit (GPU) are used for applications in which computations of
large data can be done in parallel. The general approach to performance is to divide
computations across a large number of relatively small cores. Some computationally
intensive portions are offloaded to the GPU, while the remaining computations still run
on the CPU.

Simcenter nastran GPU capabilities

AMD and NVIDIA GPU cards are supported for Simcenter nastran frequency response
(FRRD1 module) and matrix decomposition (DCMP module) computations.

• Enabling GPU computations for the FRRD1 module speeds up modal frequency
response solutions (SOL 111) when viscous or structural damping produces coupled
damping matrices. The benefit becomes significant when the number of modes is at least
5,000. See the FRRD1with GPU Performance Example in 9.3.1.

• If you have multiple GPUs, you can optionally enable GPU and DMP together for the
FRRD1 module. The gpgpu=any keyword is used to enable the GPU for FRRD1, and
the dmp=n keyword is used to specify the number of MPI threads.
Note that the maximum number of threads is equal to your number of GPU devices. All
of the GPU cards must be identical when enabling DMP. See the FRRD1 with GPU and
DMP Performance Example in 9.3.2.

• Enabling GPU computations for the DCMP module decreases the time for matrix
decomposition. The impact is more significant for sparse matrices reporting a maximum
front size larger than 30K in the .f04 file.

9.2 GPU Keyword Input Summary

The gpgpu=any keyword setting enables the GPU for both FRRD1 and DCMP module
computations. For example,
nastran.exe gpgpu=any input_file.dat

To enable GPU with DMP for a frequency response solution (FRRD1 module), you
should enable the GPU with a keyword setting such as gpgpu=any, and define the
number of MPI threads with the dmp=n keyword up to your number of GPU devices. All
of the GPU cards must be identical when enabling DMP.

100
Enabling GPU for both FRRD1 and DCMP module computations:

gpgpu
gpgpu=none Disables GPU for both FRRD1 and DCMP module computations.
gpgpu=any Enables GPU for both FRRD1 and DCMP module computations
with the first available AMD or NVIDIA GPU.

gpgpu=amd Enables GPU for both FRRD1 and DCMP module computations
with the first available AMD GPU.
gpgpu=nvidia Enables GPU for both FRRD1 and DCMP module computations
with the first available NVIDIA GPU.
gpgpu and dmp Enables GPU with DMP for the FRRD1 module computations.
keywords

Keyword cl_frrd is used to enable GPU for FRRD1 module computations only:

cl_frrd
cl_frrd=1 Enables GPU for FRRD1 module computations with the first
available AMD or NVIDIA GPU.
cl_frrd=2 Enables GPU for FRRD1 module computations with the first
available AMD GPU.

cl_frrd=3 Enables GPU for FRRD1 module computations with the first
available NVIDIA GPU.

Keyword cl_dcmp is used to enable GPU for DCMP module computations only:

cl_dcmp
cl_dcmp=1 Enables GPU for DCMP module computations with the first
available AMD or NVIDIA GPU.
cl_dcmp=2 Enables GPU for DCMP module computations with the first
available AMD GPU.
cl_dcmp=1 Enables GPU for DCMP module computations with the first
sys531=1 available NVIDIA GPU. (Both cl_dcmp=1 and sys531=1 are
required for this option)

9.3 Modal frequency response (FRRD1 module) Example

The gpgpu or the cl_frrd keyword could be used to enable the GPU for both FRRD1
module computations. For example,
nastran.exe cl_frrd=1 input_file.dat

101
The FRRD1 module will report the following in the f04 file:

16:42:31 1:31 13595.0 0.0 90.2 0.0 FREQRS 256 FRRD1 BEGN
*** USER INFORMATION MESSAGE 4157 (FRDGPU)
PARAMETERS FOR FRDGPU FOLLOW
MATRIX SIZE = 7628 NUMBER OF FREQUENCIES = 64
SYSTEM (107) = 32768 SYSTEM (573) = 1
*8** Module DMAP Matrix Cols Rows F T NzWds Density BlockT StrL NbrStr BndAvg
FRRD1 256 SCR 308 320 7628 2 4 30512 1.00000D+00 152 5204 469 7628
FRRD1 256 UHF 320 7628 2 4 30512 1.00000D+00 152 5204 469 7628

16:53:32 12:32 15161.0 1566.0 417.1 326.9 FREQRS 256 FRRD1 END

9.3.1 FRRD1 with GPU performance example

This example uses AMD Tahiti GPU (4GB), 24 core Magny-Cours.
The damping definition in the model produced coupled damping matrices.

Modes were computed up to the given frequency, where e10k = 1785 modes, e20k =
3631 modes, e30k = 5576 modes, and e40k = 7646 modes.

9.3.2 FRRD1 with GPU and DMP performance example

This example uses two AMD Tahiti GPU (4GB).
The damping definition in the model produced coupled damping matrices.

102
Modes were computed up to the given frequency, where e10k = 1785 modes, e20k =
3631 modes, e30k = 5576 modes, e40k = 7646 modes, e50k = 9787 modes and e60k =
12088 modes.

103
Appendix A

SEQP Style DMP Solutions

In the parallel processing of a finite element application, finite element model data should
be distributed to each processor. There are two choices of method for domain partitioning
in a Simcenter nastran parallel processing task: grid-based partitioning and degree of
freedom-based partitioning. The SEQP module performs grid-based partitioning, in
which p domains are created by an automatic partitioner from the grid-based graph. The
GPARTN module performs graph partitioning for the graph based on degrees of freedom
or grids, whichever is better.

The GPARTN-based method is available in more solution sequences than the SEQP-
based, and generally, the GPARTN module gives better performance than the module
SEQP. For example, the GPARTN module provides better performance if the finite
element model includes many Multi-Point Constraints (MPCs). It should be noted that
the SEQP-based method can not handle a finite element model that includes acoustic
fluid, glue elements, weld/fast elements, or modal effective mass requests. Also, grid
point weight generator output reflects only the local portion of the model (rather than
global) with the SEQP module, and thus the results may be unexpected.

The partitioning method can be selected by the nastran command keyword ‘gpart’ in the
following way.

Module Keyword
SEQP gpart = 0

GPARTN gpart = 1 default

Each module provides two different algorithms for graph partitioning. The algorithm can
be selected by the DMAP parameter oldseq.

Module Method DMAP parameter

EXTREME param, oldseq = 9

SEQP
MLV param, oldseq = 11 default

METIS param, oldseq = 10

GPARTN
MLV param, oldseq = 11 default

104
RDMODES, SOL111, SOL 112, and SOL 200 use GPARTN only.
GDMODES/HDMODES in SOL 103 and GDSTAT can use either partitioning module
(GPARTN is the default).

The following table lists the availability of domain partitioning modules for each
computational method:

SEQP GPARTN
GDSTAT X X(default)
GDMODES X X(default)
HDMODES X X(default)
RDMODES Na X(default)

In the above table, ‘X’ means the partitioning module is available. ‘na’ refers to ‘not
applicable’.

Note: Neither the module SEQP or GPARTN is used in LDSTAT, FDMODES, and
FDFREQR, because these computational methods do not require domain partitioning.

Chapter 4 includes examples for GDMODES and HDMODES with GPARTN. The
following are examples with SEQP.

A finite element plate model with 110 grid points and 100 CQUAD4 elements is
executed in parallel with SOL 103 analysis. The total number of degrees of freedom is
660. Four processors are used for the DMP run.

1. GDMODES with SEQP

GDMODES with SEQP is implemented with the nastran keyword "gpart=0".

When the domain is partitioned, the output from the SEQP module is shown as:

105
The partitioning statistics show the number of domains created, grid points, and elements
for each sub-domain.

Once the results of eigenvalue analysis on the slave processors are collected, the master
processor prints the summary of merged eigensolutions.

With the gpart=0 option, the summary of eigenvalue analysis from the master processor
is broadcasted to the slave processors, so that all processors print the same summary
information.

106


2. HDMODES with SEQP

HDMODES with SEQP is implemented with the nastran keyword "gpart=0".

In the example below, HDMODES is executed with ‘dmp=4’ and ‘nclust=2’ keywords.
Among the four processors, the first processor is the master processor, and the other
processors are the slave processors.

Note that HDMODES defines the local master and local slave processors inside of each
cluster.

nastran example dmp=4 nclust =2 gpart=0

The partitioning statistics show the information of domains for each cluster. There are
two domains for each cluster. The first domain of the cluster has 50 internal grids and 10
external grids, and the second domain has the same local size. Generally, the size of each
domain is different.

107
In this example, cluster 1 includes processors 1 and 3, and processors 2 and 4 are in
cluster 2. Processors 1 and 2 are the local masters in clusters 1 and 2, respectively. In
cluster 1, 155 eigenvalues are collected in processor 1. Similarly, in cluster 2, 53
eigenvalues are collected in processor 2. Once each local master processor collects the
results for each cluster, the master processor (processor 1) collects the information from
the local masters.

In the .f06 file, the master processor prints the summary of the eigenvalue analysis and
the list of eigenvalues. Be careful when interpreting the "number of roots found"
information in the summary. This information concerns the cluster in which the master is
included. For example, in the model below, interpret the "number of roots found"
message from the eigenvalue analysis summary as indicating that cluster 1 found 155
eigenvalues. The master processor eigenvalue table lists all 208 eigenvalues that are
merged from all of the local master processors. You should determine the number of
modes from the list of eigenvalues.

Again, be cautious when interpreting the output from slave processors. The summary of
eigenvalue analysis for slave processors is confined to the corresponding processor. In
the list of eigenvalues on the slave processors, the mode number does not represent the
global mode number.

108
 Note that, with the gpart=0, the master process broadcasts the collected output only to
the local masters. In the example below, the local slave, processor 2, prints 208
eigenvalues, while another local slave, processor 4, prints 53 eigenvalues.

109


3. GDSTAT with SEQP

GDSTAT with SEQP is implemented with the nastran keyword “gpart=0”.

Example:

The static analysis is performed for a finite element cube model that has 2197 grid points
for 1728 CHEXA elements and 6048 CQUAD4 elements. GDSTAT is implemented in a
DMP run. The module SEQP partitioned the global finite element model into four sub-
domains and a boundary. The following information in the .f06 describes the detailed
statistics of partitioning.

Each processor performs static analysis with the corresponding local domain and the
boundary, and the master processor collects the results that you requested through
communications. These outputs are printed only on the master processor.

110
It is strongly recommended to use the nastran keyword "slaveout=yes" to print out
analysis procedures from all processors. With "slaveout=yes", the f04, .f06, and .log files
contain the outputs of master and slave processors in the following format.





















Appendix B

GPGPU Library Installation

Use of AMD and NVIDIA GPGPU cards is supported in Simcenter nastran via the
OpenCL BLAS and LAPACK implementations in clBLAS. The clBLAS libraries are
used by default when the cl_frrd and cl_dcmp keywords are passed. This functionality is
available on with Windows-64 and x86_64 Linux. To install:
i. Go to the Intel Open CL website.
ii. Select the appropriate platform in the drop-down box at upper right.

111
 iii. Click Free Download.
iv. After a brief registration, the download should begin for, e.g.
intel_sdk_for_opencl_setup_6.3.0.1904.exe for Windows.
v. Install as Administrator

A second option, available only for NVIDIA GPGPUs, involves installation of the
CUDA BLAS and LAPACK implementations in cuBLAS. It is supported on both
Windows-64 and x86_64 Linux. For Linux, the necessary libraries and drivers are
available from the NVIDIA website (if not already installed) and no additional steps are
necessary. To install on Windows:
i. If no CUDA installation is present, then install CUDA version 5.0.35 using the
installer on the NVIDIA website
ii. If a newer version of the CUDA libraries is already available, then they must be in
the PATH and must be renamed to match the following:
cudart64_50_35.dll
cublas64_50_35.dll

References
[1] L. Komzsik, The Lanczos Method: Evolution and Application, 2003, SIAM.

[2] W. Gropp, E. Lusk, and A. Skjellum, Using MPI: Portable Parallel Programming with
the Message Passing Interface, 1997, MIT.

[3] Simcenter nastran Numerical Methods User’s Guide.

112