Tutorial on MPI: The

Message-Passing Interface

William Gropp

TIONAL L
NA AB
NE

OR
ARGON

ATO
RY
•U

O•
NI

AG

E
V

RS I
C

IT Y CH
OF

Mathematics and Computer Science Division

Argonne National Laboratory
Argonne, IL 60439
[email protected]
1
Course Outline
 Background on Parallel Computing
 Getting Started
 MPI Basics
 Intermediate MPI
 Tools for writing libraries
 Final comments
Thanks to Rusty Lusk for some of the material in this
tutorial.
This tutorial may be used in conjunction with
the book \Using MPI" which contains detailed
descriptions of the use of the MPI routines.
 Material that beings with this symbol is `advanced'
and may be skipped on a rst reading.

2
Background

 Parallel Computing
 Communicating with other processes
 Cooperative operations
 One-sided operations
 The MPI process

3
Parallel Computing

 Separate workers or processes

 Interact by exchanging information

4
Types of parallel computing

All use dierent data for each worker

Data-parallel Same operations on dierent
data. Also called SIMD
SPMD Same program, dierent data
MIMD Dierent programs, dierent data
SPMD and MIMD are essentially the same
because any MIMD can be made SPMD
SIMD is also equivalent, but in a less
practical sense.
MPI is primarily for SPMD/MIMD. HPF is
an example of a SIMD interface.

5
Communicating with other processes

Data must be exchanged with other workers

 Cooperative | all parties agree to
transfer data
 One sided | one worker performs
transfer of data

6
Cooperative operations

Message-passing is an approach that makes

the exchange of data cooperative.
Data must both be explicitly sent and
received.
An advantage is that any change in the
receiver's memory is made with the receiver's
participation.
Process 0 Process 1

SEND( data )

RECV( data )

7
One-sided operations
One-sided operations between parallel
processes include remote memory reads and
writes.
An advantage is that data can be accessed
without waiting for another process
Process 0 Process 1

PUT( data )

(Memory)

Process 0 Process 1

(Memory)

GET( data )

8
Class Example

Take a pad of paper. Algorithm: Initialize with the

number of neighbors you have
 Compute average of your neighbor's values and
subtract from your value. Make that your new
value.
 Repeat until done
Questions
1. How do you get values from your neighbors?
2. Which step or iteration do they correspond to?
Do you know? Do you care?
3. How do you decide when you are done?

9
Hardware models

The previous example illustrates the

hardware models by how data is exchanged
among workers.
 Distributed memory (e.g., Paragon, IBM
SPx, workstation network)
 Shared memory (e.g., SGI Power
Challenge, Cray T3D)
Either may be used with SIMD or MIMD
software models.
 All memory is distributed.

10
What is MPI?

 A message-passing library specication

{ message-passing model
{ not a compiler specication
{ not a specic product
 For parallel computers, clusters, and heterogeneous
networks
 Full-featured
 Designed to permit (unleash?) the development of
parallel software libraries
 Designed to provide access to advanced parallel
hardware for
{ end users
{ library writers
{ tool developers

11
Motivation for a New Design

 Message Passing now mature as programming

paradigm
{ well understood
{ ecient match to hardware
{ many applications
 Vendor systems not portable
 Portable systems are mostly research projects
{ incomplete
{ lack vendor support
{ not at most ecient level

12
Motivation (cont.)

Few systems oer the full range of desired features.

 modularity (for libraries)
 access to peak performance
 portability
 heterogeneity
 subgroups
 topologies
 performance measurement tools

13
The MPI Process
 Began at Williamsburg Workshop in April, 1992
 Organized at Supercomputing '92 (November)
 Followed HPF format and process
 Met every six weeks for two days
 Extensive, open email discussions
 Drafts, readings, votes
 Pre-nal draft distributed at Supercomputing '93
 Two-month public comment period
 Final version of draft in May, 1994
 Widely available now on the Web, ftp sites, netlib
(http://www.mcs.anl.gov/mpi/index.html)
 Public implementations available
 Vendor implementations coming soon

14
Who Designed MPI?
 Broad participation
 Vendors
{ IBM, Intel, TMC, Meiko, Cray, Convex, Ncube
 Library writers
{ PVM, p4, Zipcode, TCGMSG, Chameleon,
Express, Linda
 Application specialists and consultants
Companies Laboratories Universities
ARCO ANL UC Santa Barbara
Convex GMD Syracuse U
Cray Res LANL Michigan State U
IBM LLNL Oregon Grad Inst
Intel NOAA U of New Mexico
KAI NSF Miss. State U.
Meiko ORNL U of Southampton
NAG PNL U of Colorado
nCUBE Sandia Yale U
ParaSoft SDSC U of Tennessee
Shell SRC U of Maryland
TMC Western Mich U
U of Edinburgh
Cornell U.
Rice U.
U of San Francisco

15
Features of MPI

 General
{ Communicators combine context and group for
message security
{ Thread safety
 Point-to-point communication
{ Structured buers and derived datatypes,
heterogeneity
{ Modes: normal (blocking and non-blocking),
synchronous, ready (to allow access to fast
protocols), buered
 Collective
{ Both built-in and user-dened collective
operations
{ Large number of data movement routines
{ Subgroups dened directly or by topology

16
Features of MPI (cont.)

 Application-oriented process topologies

{ Built-in support for grids and graphs (uses
groups)
 Proling
{ Hooks allow users to intercept MPI calls to
install their own tools
 Environmental
{ inquiry
{ error control

17
Features not in MPI

 Non-message-passing concepts not included:

{ process management
{ remote memory transfers
{ active messages
{ threads
{ virtual shared memory
 MPI does not address these issues, but has tried to
remain compatible with these ideas (e.g. thread
safety as a goal, intercommunicators)

18
Is MPI Large or Small?

 MPI is large (125 functions)

{ MPI's extensive functionality requires many
functions
{ Number of functions not necessarily a measure
of complexity
 MPI is small (6 functions)
{ Many parallel programs can be written with just
6 basic functions.
 MPI is just right
{ One can access exibility when it is required.
{ One need not master all parts of MPI to use it.

19
Where to use MPI?

 You need a portable parallel program

 You are writing a parallel library
 You have irregular or dynamic data
relationships that do not t a data
parallel model
Where not to use MPI:
 You can use HPF or a parallel Fortran 90
 You don't need parallelism at all
 You can use libraries (which may be
written in MPI)

20
Why learn MPI?

 Portable
 Expressive
 Good way to learn about subtle issues in
parallel computing

21
Getting started
 Writing MPI programs
 Compiling and linking
 Running MPI programs
 More information
{ Using MPI by William Gropp, Ewing Lusk,
and Anthony Skjellum,
{ The LAM companion to \Using MPI..." by
Zdzislaw Meglicki
{ Designing and Building Parallel Programs by
Ian Foster.
{ A Tutorial/User's Guide for MPI by Peter
Pacheco
(ftp://math.usfca.edu/pub/MPI/mpi.guide.ps)
{ The MPI standard and other information is
available at http://www.mcs.anl.gov/mpi. Also
the source for several implementations.

22
Writing MPI programs

#include "mpi.h"
#include <stdio.h>

int main( argc, argv )

int argc;
char **argv;
{
MPI_Init( &argc, &argv );
printf( "Hello world\n" );
MPI_Finalize();
return 0;
}

23
Commentary

 #include "mpi.h"provides basic MPI

denitions and types
 MPI_Init starts MPI
 MPI_Finalize exits MPI
 Note that all non-MPI routines are local;
thus the printf run on each process

24
Compiling and linking

For simple programs, special compiler

commands can be used. For large projects,
it is best to use a standard Makele.
The MPICH implementation provides
the commands mpicc and mpif77
as well as `Makefile' examples in
`/usr/local/mpi/examples/Makefile.in'

25
Special compilation commands

The commands
mpicc -o first first.c
mpif77 -o firstf firstf.f

may be used to build simple programs when using

MPICH.
These provide special options that exploit the proling
features of MPI
-mpilog Generate log les of MPI calls
-mpitrace Trace execution of MPI calls
-mpianim Real-time animation of MPI (not available
on all systems)
There are specic to the MPICH implementation;
other implementations may provide similar commands
(e.g., mpcc and mpxlf on IBM SP2).

26
Using Makeles

The le `Makefile.in' is a template Makele.

The program (script) `mpireconfig' translates
this to a Makele for a particular system.
This allows you to use the same Makele for
a network of workstations and a massively
parallel computer, even when they use
dierent compilers, libraries, and linker
options.
mpireconfig Makefile

Note that you must have `mpireconfig' in

your PATH.

27
Sample Makele.in

##### User configurable options #####

ARCH = @ARCH@
COMM = @COMM@
INSTALL_DIR = @INSTALL_DIR@
CC = @CC@
F77 = @F77@
CLINKER = @CLINKER@
FLINKER = @FLINKER@
OPTFLAGS = @OPTFLAGS@
#
LIB_PATH = -L$(INSTALL_DIR)/lib/$(ARCH)/$(COMM)
FLIB_PATH =
@FLIB_PATH_LEADER@$(INSTALL_DIR)/lib/$(ARCH)/$(COMM)
LIB_LIST = @LIB_LIST@
#
INCLUDE_DIR = @INCLUDE_PATH@ -I$(INSTALL_DIR)/include

### End User configurable options ###

28
Sample Makele.in (con't)

CFLAGS = @CFLAGS@ $(OPTFLAGS) $(INCLUDE_DIR) -DMPI_$(ARCH)

FFLAGS = @FFLAGS@ $(INCLUDE_DIR) $(OPTFLAGS)
LIBS = $(LIB_PATH) $(LIB_LIST)
FLIBS = $(FLIB_PATH) $(LIB_LIST)
EXECS = hello

default: hello

all: $(EXECS)

hello: hello.o $(INSTALL_DIR)/include/mpi.h

$(CLINKER) $(OPTFLAGS) -o hello hello.o \
$(LIB_PATH) $(LIB_LIST) -lm

clean:
/bin/rm -f *.o *~ PI* $(EXECS)

.c.o:
$(CC) $(CFLAGS) -c $*.c
.f.o:
$(F77) $(FFLAGS) -c $*.f

29
Running MPI programs

mpirun -np 2 hello

`mpirun' is not part of the standard, but

some version of it is common with several
MPI implementations. The version shown
here is for the MPICH implementation of
MPI.
 Just as Fortran does not specify how
Fortran programs are started, MPI does not
specify how MPI programs are started.
 The option -t shows the commands that
mpirun would execute; you can use this to
nd out how mpirun starts programs on yor
system. The option -help shows all options
to mpirun.

30
Finding out about the environment

Two of the rst questions asked in a parallel

program are: How many processes are there?
and Who am I?
How many is answered with MPI_Comm_size
and who am I is answered with MPI_Comm_rank.
The rank is a number between zero and
size-1.

31
A simple program

#include "mpi.h"
#include <stdio.h>

int main( argc, argv )

int argc;
char **argv;
{
int rank, size;
MPI_Init( &argc, &argv );
MPI_Comm_rank( MPI_COMM_WORLD, &rank );
MPI_Comm_size( MPI_COMM_WORLD, &size );
printf( "Hello world! I'm %d of %d\n",
rank, size );
MPI_Finalize();
return 0;
}

32
Caveats

 These sample programs have been kept

as simple as possible by assuming that all
processes can do output. Not all parallel
systems provide this feature, and MPI
provides a way to handle this case.

33
Exercise - Getting Started

Objective: Learn how to login, write,

compile, and run a simple MPI program.
Run the \Hello world" programs. Try two
dierent parallel computers. What does the
output look like?

34
Sending and Receiving messages

Process 0 Process 1

A:
Send Recv

B:

Questions:
 To whom is data sent?
 What is sent?
 How does the receiver identify it?

35
Current Message-Passing

 A typical blocking send looks like

send( dest, type, address, length )

where
{ dest is an integer identier representing the
process to receive the message.
{ type is a nonnegative integer that the
destination can use to selectively screen
messages.
{ (address, length) describes a contiguous area in
memory containing the message to be sent.
and
 A typical global operation looks like:
broadcast( type, address, length )

 All of these specications are a good match to

hardware, easy to understand, but too in exible.

36
The Buer

Sending and receiving only a contiguous array of

bytes:
 hides the real data structure from hardware which
might be able to handle it directly
 requires pre-packing dispersed data
{ rows of a matrix stored columnwise
{ general collections of structures
 prevents communications between machines with
dierent representations (even lengths) for same
data type

37
Generalizing the Buer Description

 Specied in MPI by starting address, datatype, and

count, where datatype is:
{ elementary (all C and Fortran datatypes)
{ contiguous array of datatypes
{ strided blocks of datatypes
{ indexed array of blocks of datatypes
{ general structure
 Datatypes are constructed recursively.
 Specications of elementary datatypes allows
heterogeneous communication.
 Elimination of length in favor of count is clearer.
 Specifying application-oriented layout of data
allows maximal use of special hardware.

38
Generalizing the Type

 A single type eld is too constraining. Often

overloaded to provide needed exibility.
 Problems:
{ under user control
{ wild cards allowed (MPI_ANY_TAG)
{ library use con icts with user and with other
libraries

39
Sample Program using Library Calls

Sub1 and Sub2 are from dierent libraries.

Sub1();
Sub2();

Sub1a and Sub1b are from the same library

Sub1a();
Sub2();
Sub1b();

Thanks to Marc Snir for the following four examples

40
Correct Execution of Library Calls

Process 0 Process 1 Process 2

recv(any) send(1)
Sub1

recv(any) send(0)

recv(1) send(0)
Sub2
recv(2) send(1)

send(2) recv(0)

41
Incorrect Execution of Library Calls

Process 0 Process 1 Process 2

recv(any) send(1)
Sub1

recv(any) send(0)

recv(1) send(0)

Sub2 recv(2) send(1)

send(2) recv(0)

42
Correct Execution of Library Calls with Pending
Communcication

Process 0 Process 1 Process 2

recv(any) send(1)
Sub1a

send(0)

recv(2) send(0)

Sub2 send(2) recv(1)

send(1) recv(0)

recv(any)
Sub1b

43
Incorrect Execution of Library Calls with Pending
Communication

Process 0 Process 1 Process 2

recv(any) send(1)
Sub1a

send(0)

recv(2) send(0)

Sub2 send(2) recv(1)

send(1) recv(0)

recv(any)

Sub1b

44
Solution to the type problem

 A separate communication context for each family

of messages, used for queueing and matching.
(This has often been simulated in the past by
overloading the tag eld.)
 No wild cards allowed, for security
 Allocated by the system, for security
 Types (tags, in MPI) retained for normal use (wild
cards OK)

45
Delimiting Scope of Communication

 Separate groups of processes working on

subproblems
{ Merging of process name space interferes with
modularity
{ \Local" process identiers desirable
 Parallel invocation of parallel libraries
{ Messages from application must be kept
separate from messages internal to library.
{ Knowledge of library message types interferes
with modularity.
{ Synchronizing before and after library calls is
undesirable.

46
Generalizing the Process Identier

 Collective operations typically operated on all

processes (although some systems provide
subgroups).
 This is too restrictive (e.g., need minimum over a
column or a sum across a row, of processes)
 MPI provides groups of processes
{ initial \all" group
{ group management routines (build, delete
groups)
 All communication (not just collective operations)
takes place in groups.
 A group and a context are combined in a
communicator.
 Source/destination in send/receive operations refer
to rank in group associated with a given
communicator. MPI_ANY_SOURCE permitted in a
receive.

47
MPI Basic Send/Receive

Thus the basic (blocking) send has become:

MPI_Send( start, count, datatype, dest, tag,
comm )

and the receive:

MPI_Recv(start, count, datatype, source, tag,
comm, status)

The source, tag, and count of the message actually

received can be retrieved from status.
Two simple collective operations:
MPI_Bcast(start, count, datatype, root, comm)
MPI_Reduce(start, result, count, datatype,
operation, root, comm)

48
Getting information about a message

MPI_Status status;
MPI_Recv( ..., &status );
... status.MPI_TAG;
... status.MPI_SOURCE;
MPI_Get_count( &status, datatype, &count );

MPI_TAG and MPI_SOURCE primarily of use when

MPI_ANY_TAG and/or MPI_ANY_SOURCE in the receive.
MPI_Get_count may be used to determine how much
data of a particular type was received.

49
Simple Fortran example
program main
include 'mpif.h'

integer rank, size, to, from, tag, count, i, ierr

integer src, dest
integer st_source, st_tag, st_count
integer status(MPI_STATUS_SIZE)
double precision data(100)

call MPI_INIT( ierr )

call MPI_COMM_RANK( MPI_COMM_WORLD, rank, ierr )
call MPI_COMM_SIZE( MPI_COMM_WORLD, size, ierr )
print *, 'Process ', rank, ' of ', size, ' is alive'
dest = size - 1
src = 0
C
if (rank .eq. src) then
to = dest
count = 10
tag = 2001
do 10 i=1, 10
10 data(i) = i
call MPI_SEND( data, count, MPI_DOUBLE_PRECISION, to,
+ tag, MPI_COMM_WORLD, ierr )
else if (rank .eq. dest) then
tag = MPI_ANY_TAG
count = 10
from = MPI_ANY_SOURCE
call MPI_RECV(data, count, MPI_DOUBLE_PRECISION, from,
+ tag, MPI_COMM_WORLD, status, ierr )

50
Simple Fortran example (cont.)

call MPI_GET_COUNT( status, MPI_DOUBLE_PRECISION,

+ st_count, ierr )
st_source = status(MPI_SOURCE)
st_tag = status(MPI_TAG)
C
print *, 'Status info: source = ', st_source,
+ ' tag = ', st_tag, ' count = ', st_count
print *, rank, ' received', (data(i),i=1,10)
endif

call MPI_FINALIZE( ierr )

end

51
Six Function MPI

MPI is very simple. These six functions allow

you to write many programs:
MPI Init
MPI Finalize
MPI Comm size
MPI Comm rank
MPI Send
MPI Recv

52
A taste of things to come
The following examples show a C and
Fortran version of the same program.
This program computes PI (with a very
simple method) but does not use MPI_Send
and MPI_Recv. Instead, it uses collective
operations to send data to and from all of
the running processes. This gives a dierent
six-function MPI set:
MPI Init
MPI Finalize
MPI Comm size
MPI Comm rank
MPI Bcast
MPI Reduce
53
Broadcast and Reduction

The routine MPI_Bcast sends data from one

process to all others.
The routine MPI_Reduce combines data from
all processes (by adding them in this case),
and returning the result to a single process.

54
Fortran example: PI

program main

include "mpif.h"

double precision PI25DT

parameter (PI25DT = 3.141592653589793238462643d0)

double precision mypi, pi, h, sum, x, f, a

integer n, myid, numprocs, i, rc
c function to integrate
f(a) = 4.d0 / (1.d0 + a*a)

call MPI_INIT( ierr )

call MPI_COMM_RANK( MPI_COMM_WORLD, myid, ierr )
call MPI_COMM_SIZE( MPI_COMM_WORLD, numprocs, ierr )

10 if ( myid .eq. 0 ) then

write(6,98)
98 format('Enter the number of intervals: (0 quits)')
read(5,99) n
99 format(i10)
endif

call MPI_BCAST(n,1,MPI_INTEGER,0,MPI_COMM_WORLD,ierr)

55
Fortran example (cont.)
c check for quit signal
if ( n .le. 0 ) goto 30

c calculate the interval size

h = 1.0d0/n

sum = 0.0d0
do 20 i = myid+1, n, numprocs
x = h * (dble(i) - 0.5d0)
sum = sum + f(x)
20 continue
mypi = h * sum

c collect all the partial sums

call MPI_REDUCE(mypi,pi,1,MPI_DOUBLE_PRECISION,MPI_SUM,0,
$ MPI_COMM_WORLD,ierr)

c node 0 prints the answer.

if (myid .eq. 0) then
write(6, 97) pi, abs(pi - PI25DT)
97 format(' pi is approximately: ', F18.16,
+ ' Error is: ', F18.16)
endif

goto 10

30 call MPI_FINALIZE(rc)
stop
end

56
C example: PI

#include "mpi.h"
#include <math.h>

int main(argc,argv)
int argc;
char *argv[];
{
int done = 0, n, myid, numprocs, i, rc;
double PI25DT = 3.141592653589793238462643;
double mypi, pi, h, sum, x, a;

MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&numprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&myid);

57
C example (cont.)

while (!done)
{
if (myid == 0) {
printf("Enter the number of intervals: (0 quits) ");
scanf("%d",&n);
}
MPI_Bcast(&n, 1, MPI_INT, 0, MPI_COMM_WORLD);
if (n == 0) break;

h = 1.0 / (double) n;
sum = 0.0;
for (i = myid + 1; i <= n; i += numprocs) {
x = h * ((double)i - 0.5);
sum += 4.0 / (1.0 + x*x);
}
mypi = h * sum;

MPI_Reduce(&mypi, &pi, 1, MPI_DOUBLE, MPI_SUM, 0,

MPI_COMM_WORLD);

if (myid == 0)
printf("pi is approximately %.16f, Error is %.16f\n",
pi, fabs(pi - PI25DT));
}
MPI_Finalize();
}

58
Exercise - PI

Objective: Experiment with send/receive

Run either program for PI. Write new
versions that replace the calls to MPI_Bcast
and MPI_Reduce with MPI_Send and MPI_Recv.
 The MPI broadcast and reduce operations
use at most log p send and receive operations
on each process where p is the size of
MPI COMM WORLD. How many operations do
your versions use?

59
Exercise - Ring

Objective: Experiment with send/receive

Write a program to send a message around a
ring of processors. That is, processor 0 sends
to processor 1, who sends to processor 2,
etc. The last processor returns the message
to processor 0.
 You can use the routine MPI Wtime to time
code in MPI. The statement
t = MPI Wtime();

returns the time as a double (DOUBLE

PRECISION in Fortran).

60
Topologies

MPI provides routines to provide structure to

collections of processes
This helps to answer the question:
Who are my neighbors?

61
Cartesian Topologies

A Cartesian topology is a mesh

Example of 3  4 Cartesian mesh with arrows
pointing at the right neighbors:
(0,2) (1,2) (2,2) (3,2)

(0,1) (1,1) (2,1) (3,1)

(0,0) (1,0) (2,0) (3,0)

62
Dening a Cartesian Topology

The routine MPI_Cart_create creates a Cartesian

decomposition of the processes, with the number of
dimensions given by the ndim argument.
dims(1) = 4
dims(2) = 3
periods(1) = .false.
periods(2) = .false.
reorder = .true.
ndim = 2
call MPI_CART_CREATE( MPI_COMM_WORLD, ndim, dims,
$ periods, reorder, comm2d, ierr )

63
Finding neighbors

MPI_Cart_create creates a new communicator with the

same processes as the input communicator, but with
the specied topology.
The question, Who are my neighbors, can now be
answered with MPI_Cart_shift:
call MPI_CART_SHIFT( comm2d, 0, 1,
nbrleft, nbrright, ierr )
call MPI_CART_SHIFT( comm2d, 1, 1,
nbrbottom, nbrtop, ierr )

The values returned are the ranks, in the

communicator comm2d, of the neighbors shifted by 1
in the two dimensions.

64
Who am I?

Can be answered with

integer coords(2)
call MPI_COMM_RANK( comm1d, myrank, ierr )
call MPI_CART_COORDS( comm1d, myrank, 2,
$ coords, ierr )

Returns the Cartesian coordinates of the calling

process in coords.

65
Partitioning

When creating a Cartesian topology, one question is

\What is a good choice for the decomposition of the
processors?"
This question can be answered with MPI_Dims_create:
integer dims(2)
dims(1) = 0
dims(2) = 0
call MPI_COMM_SIZE( MPI_COMM_WORLD, size, ierr )
call MPI_DIMS_CREATE( size, 2, dims, ierr )

66
Other Topology Routines

MPI contains routines to translate between

Cartesian coordinates and ranks in a
communicator, and to access the properties
of a Cartesian topology.
The routine MPI_Graph_create allows the
creation of a general graph topology.

67
Why are these routines in MPI?
In many parallel computer interconnects,
some processors are closer to than
others. These routines allow the MPI
implementation to provide an ordering of
processes in a topology that makes logical
neighbors close in the physical interconnect.
 Some parallel programmers may remember
hypercubes and the eort that went into
assigning nodes in a mesh to processors
in a hypercube through the use of Grey
codes. Many new systems have dierent
interconnects; ones with multiple paths
may have notions of near neighbors that
changes with time. These routines free
the programmer from many of these
considerations. The reorder argument is
used to request the best ordering.
68
The periods argument

Who are my neighbors if I am at the edge of

a Cartesian Mesh?
?

69
Periodic Grids

Specify this in MPI_Cart_create with

dims(1) = 4
dims(2) = 3
periods(1) = .TRUE.
periods(2) = .TRUE.
reorder = .true.
ndim = 2
call MPI_CART_CREATE( MPI_COMM_WORLD, ndim, dims,
$ periods, reorder, comm2d, ierr )

70
Nonperiodic Grids

In the nonperiodic case, a neighbor may

not exist. This is indicated by a rank of
MPI_PROC_NULL.

This rank may be used in send and receive

calls in MPI. The action in both cases is as if
the call was not made.

71
Collective Communications in MPI

 Communication is coordinated among a group of

processes.
 Groups can be constructed \by hand" with MPI
group-manipulation routines or by using MPI
topology-denition routines.
 Message tags are not used. Dierent
communicators are used instead.
 No non-blocking collective operations.
 Three classes of collective operations:
{ synchronization
{ data movement
{ collective computation

72
Synchronization

 MPI_Barrier(comm)
 Function blocks untill all processes in
comm call it.

73
Available Collective Patterns
P0 A P0 A

P1 Broadcast P1 A

P2 P2 A

P3 P3 A

P0 A B C D P0 A
Scatter
P1 P1 B

P2 P2 C
Gather
P3 P3 D

P0 A P0 A B C D

P1 B All gather P1 A B C D

P2 C P2 A B C D

P3 D P3 A B C D

P0 A0 A1 A2 A3 P0 A0 B0 C0 D0

P1 B0 B1 B2 B3 All to All P1 A1 B1 C1 D1

P2 C0 C1 C2 C3 P2 A2 B2 C2 D2

P3 D0 D1 D2 D3 P3 A3 B3 C3 D3

Schematic representation of collective data

movement in MPI

74
Available Collective Computation Patterns

P0 A P0 ABCD

P1 B Reduce P1

P2 C P2

P3 D P3

P0 A P0 A

P1 B P1 AB
Scan
P2 C P2 ABC

P3 D P3 ABCD

Schematic representation of collective data

movement in MPI

75
MPI Collective Routines

 Many routines:
Allgather Allgatherv Allreduce
Alltoall Alltoallv Bcast
Gather Gatherv Reduce
ReduceScatter Scan Scatter
Scatterv
 All versions deliver results to all participating
processes.
 V versions allow the chunks to have dierent sizes.
 Allreduce, Reduce, ReduceScatter, and Scan take
both built-in and user-dened combination
functions.

76
Built-in Collective Computation Operations

MPI Name Operation

MPI MAX Maximum
MPI MIN Minimum
MPI PROD Product
MPI SUM Sum
MPI LAND Logical and
MPI LOR Logical or
MPI LXOR Logical exclusive or (xor)
MPI BAND Bitwise and
MPI BOR Bitwise or
MPI BXOR Bitwise xor
MPI MAXLOC Maximum value and location
MPI MINLOC Minimum value and location

77
Dening Your Own Collective Operations

MPI_Op_create(user_function, commute, op)

MPI_Op_free(op)

user_function(invec, inoutvec, len, datatype)

The user function should perform

inoutvec[i] = invec[i] op inoutvec[i];

for i from 0 to len-1.

user_function can be non-commutative (e.g., matrix
multiply).

78
Sample user function

For example, to create an operation that has the

same eect as MPI_SUM on Fortran double precision
values, use
subroutine myfunc( invec, inoutvec, len, datatype )
integer len, datatype
double precision invec(len), inoutvec(len)
integer i
do 10 i=1,len
10 inoutvec(i) = invec(i) + inoutvec(i)
return
end

To use, just
integer myop
call MPI_Op_create( myfunc, .true., myop, ierr )
call MPI_Reduce( a, b, 1, MPI_DOUBLE_PRECISON, myop, ... )

The routine MPI_Op_free destroys user-functions when

they are no longer needed.

79
Dening groups

All MPI communication is relative to a

communicator which contains a context
and a group. The group is just a set of
processes.

80
Subdividing a communicator

The easiest way to create communicators with new

groups is with MPI_COMM_SPLIT.
For example, to form groups of rows of processes
Column
0 1 2 3 4
0

Row 1

use
MPI_Comm_split( oldcomm, row, 0, &newcomm );

To maintain the order by rank, use

MPI_Comm_rank( oldcomm, &rank );
MPI_Comm_split( oldcomm, row, rank, &newcomm );

81
Subdividing (con't)

Similarly, to form groups of columns,

Column
0 1 2 3 4
0

Row 1

use
MPI_Comm_split( oldcomm, column, 0, &newcomm2 );

To maintain the order by rank, use

MPI_Comm_rank( oldcomm, &rank );
MPI_Comm_split( oldcomm, column, rank, &newcomm2 );

82
Manipulating Groups

Another way to create a communicator with specic

members is to use MPI_Comm_create.
MPI_Comm_create( oldcomm, group, &newcomm );

The group can be created in many ways:

83
Creating Groups

All group creation routines create a group by

specifying the members to take from an existing
group.
 MPI_Group_incl species specic members
 MPI_Group_excl excludes specic members
 and MPI_Group_range_excl
MPI_Group_range_incl
use ranges of members
 MPI_Group_union and MPI_Group_intersection
creates a new group from two existing groups.
To get an existing group, use
MPI_Comm_group( oldcomm, &group );

Free a group with

MPI_Group_free( &group );

84
Buering issues

Where does data go when you send it? One

possibility is:
Process 1 Process 2
A:

Local Buffer

Local Buffer

The Network B:

85
Better buering

This is not very ecient. There are three

copies in addition to the exchange of data
between processes. We prefer
Process 1 Process 2
A:

B:

But this requires that either that MPI_Send

not return until the data has been delivered
or that we allow a send operation to return
before completing the transfer. In this case,
we need to test for completion later.

86
Blocking and Non-Blocking communication

 So far we have used blocking communication:

{ MPI Send does not complete until buer is empty
(available for reuse).
{ MPI Recv does not complete until buer is full
(available for use).
 Simple, but can be \unsafe":
Process 0 Process 1
Send(1) Send(0)
Recv(1) Recv(0)
Completion depends in general on size of message
and amount of system buering.
 Send works for small enough messages but fails
when messages get too large. Too large ranges from
zero bytes to 100's of Megabytes.

87
Some Solutions to the \Unsafe" Problem

 Order the operations more carefully:

Process 0 Process 1
Send(1) Recv(0)
Recv(1) Send(0)
 Supply receive buer at same time as send, with
MPI Sendrecv:

Process 0 Process 1
Sendrecv(1) Sendrecv(0)
 Use non-blocking operations:
Process 0 Process 1
Isend(1) Isend(0)
Irecv(1) Irecv(0)
Waitall Waitall
 Use MPI_Bsend

88
MPI's Non-Blocking Operations

Non-blocking operations return (immediately)

\request handles" that can be waited on and queried:
 MPI Isend(start, count, datatype, dest, tag, comm,
request)

 MPI Irecv(start, count, datatype, dest, tag, comm,

request)

 MPI Wait(request, status)

One can also test without waiting: MPI_Test( request,

flag, status)

89
Multiple completions

It is often desirable to wait on multiple requests. An

example is a master/slave program, where the master
waits for one or more slaves to send it a message.
 MPI Waitall(count, array of requests,
array of statuses)

 MPI Waitany(count, array of requests, index,

status)

 MPI Waitsome(incount, array of requests, outcount,

array of indices, array of statuses)

There are corresponding versions of test for each of

these.
 The MPI WAITSOME and MPI TESTSOME may be used to
implement master/slave algorithms that provide fair
access to the master by the slaves.

90
Fairness
What happens with this program:
#include "mpi.h"
#include <stdio.h>
int main(argc, argv)
int argc;
char **argv;
{
int rank, size, i, buf[1];
MPI_Status status;

MPI_Init( &argc, &argv );

MPI_Comm_rank( MPI_COMM_WORLD, &rank );
MPI_Comm_size( MPI_COMM_WORLD, &size );
if (rank == 0) {
for (i=0; i<100*(size-1); i++) {
MPI_Recv( buf, 1, MPI_INT, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, &status );
printf( "Msg from %d with tag %d\n",
status.MPI_SOURCE, status.MPI_TAG );
}
}
else {
for (i=0; i<100; i++)
MPI_Send( buf, 1, MPI_INT, 0, i, MPI_COMM_WORLD );
}
MPI_Finalize();
return 0;
}

91
Fairness in message-passing

An parallel algorithm is fair if no process

is eectively ignored. In the preceeding
program, processes with low rank (like
process zero) may be the only one whose
messages are received.
MPI makes no guarentees about fairness.
However, MPI makes it possible to write
ecient, fair programs.

92
Providing Fairness

One alternative is
#define large 128
MPI_Request requests[large];
MPI_Status statuses[large];
int indices[large];
int buf[large];
for (i=1; i<size; i++)
MPI_Irecv( buf+i, 1, MPI_INT, i,
MPI_ANY_TAG, MPI_COMM_WORLD, &requests[i-1] );
while(not done) {
MPI_Waitsome( size-1, requests, &ndone, indices, statuses );
for (i=0; i<ndone; i++) {
j = indices[i];
printf( "Msg from %d with tag %d\n",
statuses[i].MPI_SOURCE,
statuses[i].MPI_TAG );
MPI_Irecv( buf+j, 1, MPI_INT, j,
MPI_ANY_TAG, MPI_COMM_WORLD, &requests[j] );
}
}

93
Providing Fairness (Fortran)

One alternative is
parameter( large = 128 )
integer requests(large);
integer statuses(MPI_STATUS_SIZE,large);
integer indices(large);
integer buf(large);
logical done
do 10 i = 1,size-1
10 call MPI_Irecv( buf(i), 1, MPI_INTEGER, i,
* MPI_ANY_TAG, MPI_COMM_WORLD, requests(i), ierr )
20 if (.not. done) then
call MPI_Waitsome( size-1, requests, ndone,
indices, statuses, ierr )
do 30 i=1, ndone
j = indices(i)
print *, 'Msg from ', statuses(MPI_SOURCE,i), ' with tag',
* statuses(MPI_TAG,i)
call MPI_Irecv( buf(j), 1, MPI_INTEGER, j,
MPI_ANY_TAG, MPI_COMM_WORLD, requests(j), ierr )
done = ...
30 continue
goto 20
endif

94
Exercise - Fairness

Objective: Use nonblocking communications

Complete the program fragment on
\providing fairness". Make sure that you
leave no uncompleted requests. How would
you test your program?

95
More on nonblocking communication

In applications where the time to send data between

processes is large, it is often helpful to cause
communication and computation to overlap. This can
easily be done with MPI's non-blocking routines.
For example, in a 2-D nite dierence mesh, moving
data needed for the boundaries can be done at the
same time as computation on the interior.
MPI_Irecv( ... each ghost edge ... );
MPI_Isend( ... data for each ghost edge ... );
... compute on interior
while (still some uncompleted requests) {
MPI_Waitany( ... requests ... )
if (request is a receive)
... compute on that edge ...
}

Note that we call MPI_Waitany several times. This

exploits the fact that after a request is satised, it
is set to MPI_REQUEST_NULL, and that this is a valid
request object to the wait and test routines.

96
Communication Modes

MPI provides mulitple modes for sending messages:

 Synchronous mode (MPI Ssend): the send does not
complete until a matching receive has begun.
(Unsafe programs become incorrect and usually
deadlock within an MPI_Ssend.)
 Buered mode (MPI Bsend): the user supplies the
buer to system for its use. (User supplies enough
memory to make unsafe program safe).
 Ready mode (MPI Rsend): user guarantees that
matching receive has been posted.
{ allows access to fast protocols
{ undened behavior if the matching receive is not
posted
Non-blocking versions:
MPI Issend, MPI Irsend, MPI Ibsend

Note that an MPI_Recv may receive messages sent with

any send mode.

97
Buered Send
MPI provides a send routine that may be used when
MPI_Isend is awkward to use (e.g., lots of small
messages).
MPI_Bsend makes use of a user-provided buer to save
any messages that can not be immediately sent.
int bufsize;
char *buf = malloc(bufsize);
MPI_Buffer_attach( buf, bufsize );
...
MPI_Bsend( ... same as MPI_Send ... );
...
MPI_Buffer_detach( &buf, &bufsize );

The MPI_Buffer_detach call does not complete until all

messages are sent.
 The performance of MPI Bsend depends on the
implementation of MPI and may also depend on
the size of the message. For example, making a
message one byte longer may cause a signicant drop
in performance.

98
Reusing the same buer

Consider a loop
MPI_Buffer_attach( buf, bufsize );
while (!done) {
...
MPI_Bsend( ... );
}

where the buf is large enough to hold the message in

the MPI_Bsend. This code may fail because the
{
void *buf; int bufsize;
MPI_Buffer_detach( &buf, &bufsize );
MPI_Buffer_attach( buf, bufsize );
}

99
Other Point-to-Point Features

 MPI_SENDRECV, MPI_SENDRECV_REPLACE
 MPI_CANCEL
 Persistent communication requests

100
Datatypes and Heterogenity

MPI datatypes have two main purposes

 Heterogenity | parallel programs
between dierent processors
 Noncontiguous data | structures,
vectors with non-unit stride, etc.
Basic datatype, corresponding to the
underlying language, are predened.
The user can construct new datatypes at run
time; these are called derived datatypes.

101
Datatypes in MPI

Elementary: Language-dened types (e.g.,

MPI_INT or MPI_DOUBLE_PRECISION )
Vector: Separated by constant \stride"
Contiguous: Vector with stride of one
Hvector: Vector, with stride in bytes
Indexed: Array of indices (for
scatter/gather)
Hindexed: Indexed, with indices in bytes
Struct: General mixed types (for C structs
etc.)

102
Basic Datatypes (Fortran)

MPI datatype Fortran datatype

MPI_INTEGER INTEGER
MPI_REAL REAL
MPI_DOUBLE_PRECISION DOUBLE PRECISION
MPI_COMPLEX COMPLEX
MPI_LOGICAL LOGICAL
MPI_CHARACTER CHARACTER(1)
MPI_BYTE
MPI_PACKED

103
Basic Datatypes (C)

MPI datatype C datatype

MPI_CHAR signed char
MPI_SHORT signed short int
MPI_INT signed int
MPI_LONG signed long int
MPI_UNSIGNED_CHAR unsigned char
MPI_UNSIGNED_SHORT unsigned short int
MPI_UNSIGNED unsigned int
MPI_UNSIGNED_LONG unsigned long int
MPI_FLOAT float
MPI_DOUBLE double
MPI_LONG_DOUBLE long double
MPI_BYTE
MPI_PACKED

104
Vectors

29 30 31 32 33 34 35

22 23 24 25 26 27 28

15 16 17 18 19 20 21

8 9 10 11 12 13 14

1 2 3 4 5 6 7

To specify this row (in C order), we can use

MPI_Type_vector( count, blocklen, stride, oldtype,
&newtype );
MPI_Type_commit( &newtype );

The exact code for this is

MPI_Type_vector( 5, 1, 7, MPI_DOUBLE, &newtype );
MPI_Type_commit( &newtype );

105
Structures

Structures are described by arrays of

 number of elements (array_of_len)
 displacement or location (array_of_displs)
 datatype (array_of_types)
MPI_Type_structure( count, array_of_len,
array_of_displs,
array_of_types, &newtype );

106
Example: Structures

struct {
char display[50]; /* Name of display */
int maxiter; /* max # of iterations */
double xmin, ymin; /* lower left corner of rectangle */
double xmax, ymax; /* upper right corner */
int width; /* of display in pixels */
int height; /* of display in pixels */
} cmdline;

/* set up 4 blocks */
int blockcounts[4] = {50,1,4,2};
MPI_Datatype types[4];
MPI_Aint displs[4];
MPI_Datatype cmdtype;

/* initialize types and displs with addresses of items */

MPI_Address( &cmdline.display, &displs[0] );
MPI_Address( &cmdline.maxiter, &displs[1] );
MPI_Address( &cmdline.xmin, &displs[2] );
MPI_Address( &cmdline.width, &displs[3] );
types[0] = MPI_CHAR;
types[1] = MPI_INT;
types[2] = MPI_DOUBLE;
types[3] = MPI_INT;
for (i = 3; i >= 0; i--)
displs[i] -= displs[0];
MPI_Type_struct( 4, blockcounts, displs, types, &cmdtype );
MPI_Type_commit( &cmdtype );

107
Strides

The extent of a datatype is (normally) the

distance between the rst and last member.
Memory locations specified by datatype

EXTENT
LB UB

You can set an articial extent by using

MPI_UB and MPI_LB in MPI_Type_struct.

108
Vectors revisited

This code creates a datatype for an arbitrary

number of element in a row of an array
stored in Fortran order (column rst).
int blens[2], displs[2];
MPI_Datatype types[2], rowtype;
blens[0] = 1;
blens[1] = 1;
displs[0] = 0;
displs[1] = number_in_column * sizeof(double);
types[0] = MPI_DOUBLE;
types[1] = MPI_UB;
MPI_Type_struct( 2, blens, displs, types, &rowtype );
MPI_Type_commit( &rowtype );

To send n elements, you can use

MPI_Send( buf, n, rowtype, ... );

109
Structures revisited

When sending an array of a structure, it is important

to ensure that MPI and the C compiler have the
same value for the size of each structure. The most
portable way to do this is to add an MPI_UB to the
structure denition for the end of the structure. In
the previous example, this is
/* initialize types and displs with addresses of items */
MPI_Address( &cmdline.display, &displs[0] );
MPI_Address( &cmdline.maxiter, &displs[1] );
MPI_Address( &cmdline.xmin, &displs[2] );
MPI_Address( &cmdline.width, &displs[3] );
MPI_Address( &cmdline+1, &displs[4] );
types[0] = MPI_CHAR;
types[1] = MPI_INT;
types[2] = MPI_DOUBLE;
types[3] = MPI_INT;
types[4] = MPI_UB;
for (i = 4; i >= 0; i--)
displs[i] -= displs[0];
MPI_Type_struct( 5, blockcounts, displs, types, &cmdtype );
MPI_Type_commit( &cmdtype );

110
Interleaving data

By moving the UB inside the data, you can

interleave data.
Consider the matrix
0 8 16 24 32
1 9 17 25 33
2 10 18 26 34
3 11 19 27 35
4 12 20 28 36
5 13 21 29 37
6 14 22 30 38
7 15 23 31 39

We wish to send 0-3,8-11,16-19, and 24-27

to process 0, 4-7,12-15,20-23, and 28-31 to
process 1, etc. How can we do this with
MPI_Scatterv?

111
An interleaved datatype

MPI_Type_vector( 4, 4, 8, MPI_DOUBLE, &vec );

denes a block of this matrix.

blens[0] = 1; blens[1] = 1;
types[0] = vec; types[1] = MPI_UB;
displs[0] = 0; displs[1] = sizeof(double);
MPI_Type_struct( 2, blens, displs, types, &block );

denes a block whose extent is just 1 entries.

112
Scattering a Matrix

We set the displacements for each block as the

location of the rst element in the block. This works
because MPI_Scatterv uses the extents to determine
the start of each piece to send.
scdispls[0] = 0;
scdispls[1] = 4;
scdispls[2] = 32;
scdispls[3] = 36;
MPI_Scatterv( sendbuf, sendcounts, scdispls, block,
recvbuf, nx * ny, MPI_DOUBLE, 0,
MPI_COMM_WORLD );

 How would use use the topology routines to make

this more general?

113
Exercises - datatypes
Objective: Learn about datatypes
1. Write a program to send rows of a matrix (stored
in column-major form) to the other processors.
Let processor 0 have the entire matrix, which has
as many rows as processors.
Processor 0 sends row i to processor i.
Processor i reads that row into a local array that
holds only that row. That is, processor 0 has a
matrix A(N; M ) while the other processors have a
row B(M ).
(a) Write the program to handle the case where
the matrix is square.
(b) Write the program to handle a number of
columns read from the terminal.
C programmers may send columns of a matrix
stored in row-major form if they prefer.
If you have time, try one of the following. If you
don't have time, think about how you would
program these.
2. Write a program to transpose a matrix, where
each processor has a part of the matrix. Use
topologies to dene a 2-Dimensional partitioning
114
 of the matrix across the processors, and assume
that all processors have the same size submatrix.
(a) Use MPI_Send and MPI_Recv to send the block,
the transpose the block.
(b) Use MPI_Sendrecv instead.
(c) Create a datatype that allows you to receive
the block already transposed.
3. Write a program to send the "ghostpoints" of a
2-Dimensional mesh to the neighboring
processors. Assume that each processor has the
same size subblock.
(a) Use topologies to nd the neighbors
(b) Dene a datatype for the \rows"
(c) Use MPI_Sendrecv or MPI_IRecv and MPI_Send
with MPI_Waitall.
(d) Use MPI_Isend and MPI_Irecv to start the
communication, do some computation on the
interior, and then use MPI_Waitany to process
the boundaries as they arrive
The same approach works for general
datastructures, such as unstructured meshes.
4. Do 3, but for 3-Dimensional meshes. You will
need MPI_Type_Hvector.
Tools for writing libraries

MPI is specically designed to make it easier

to write message-passing libraries
 Communicators solve tag/source
wild-card problem
 Attributes provide a way to attach
information to a communicator

115
Private communicators

One of the rst thing that a library should

normally do is create private communicator.
This allows the library to send and receive
messages that are known only to the library.
MPI_Comm_dup( old_comm, &new_comm );

116
Attributes

Attributes are data that can be attached to

one or more communicators.
Attributes are referenced by keyval. Keyvals
are created with MPI_KEYVAL_CREATE.
Attributes are attached to a communicator
with MPI_Attr_put and their values accessed
by MPI_Attr_get.
 Operations are dened for what happens
to an attribute when it is copied (by creating
one communicator from another) or deleted
(by deleting a communicator) when the
keyval is created.

117
What is an attribute?

In C, an attribute is a pointer of type void *.

You must allocate storage for the attribute
to point to (make sure that you don't use
the address of a local variable).
In Fortran, it is a single INTEGER.

118
Examples of using attributes

 Forcing sequential operation

 Managing tags

119
Sequential Sections
#include "mpi.h"
#include <stdlib.h>

static int MPE_Seq_keyval = MPI_KEYVAL_INVALID;

/*@
MPE_Seq_begin - Begins a sequential section of code.

Input Parameters:
. comm - Communicator to sequentialize.
. ng - Number in group. This many processes are allowed
to execute
at the same time. Usually one.

@*/
void MPE_Seq_begin( comm, ng )
MPI_Comm comm;
int ng;
{
int lidx, np;
int flag;
MPI_Comm local_comm;
MPI_Status status;

/* Get the private communicator for the sequential

operations */
if (MPE_Seq_keyval == MPI_KEYVAL_INVALID) {
MPI_Keyval_create( MPI_NULL_COPY_FN,
MPI_NULL_DELETE_FN,
&MPE_Seq_keyval, NULL );
}

120
Sequential Sections II

MPI_Attr_get( comm, MPE_Seq_keyval, (void *)&local_comm,

&flag );
if (!flag) {
/* This expects a communicator to be a pointer */
MPI_Comm_dup( comm, &local_comm );
MPI_Attr_put( comm, MPE_Seq_keyval,
(void *)local_comm );
}
MPI_Comm_rank( comm, &lidx );
MPI_Comm_size( comm, &np );
if (lidx != 0) {
MPI_Recv( NULL, 0, MPI_INT, lidx-1, 0, local_comm,
&status );
}
/* Send to the next process in the group unless we
are the last process in the processor set */
if ( (lidx % ng) < ng - 1 && lidx != np - 1) {
MPI_Send( NULL, 0, MPI_INT, lidx + 1, 0, local_comm );
}
}

121
Sequential Sections III
/*@
MPE_Seq_end - Ends a sequential section of code.
Input Parameters:
. comm - Communicator to sequentialize.
. ng - Number in group.
@*/
void MPE_Seq_end( comm, ng )
MPI_Comm comm;
int ng;
{
int lidx, np, flag;
MPI_Status status;
MPI_Comm local_comm;

MPI_Comm_rank( comm, &lidx );

MPI_Comm_size( comm, &np );
MPI_Attr_get( comm, MPE_Seq_keyval, (void *)&local_comm,
&flag );
if (!flag)
MPI_Abort( comm, MPI_ERR_UNKNOWN );
/* Send to the first process in the next group OR to the
first process
in the processor set */
if ( (lidx % ng) == ng - 1 || lidx == np - 1) {
MPI_Send( NULL, 0, MPI_INT, (lidx + 1) % np, 0,
local_comm );
}
if (lidx == 0) {
MPI_Recv( NULL, 0, MPI_INT, np-1, 0, local_comm,
&status );
}
}

122
Comments on sequential sections

 Note use of MPI_KEYVAL_INVALID to

determine to create a keyval
 Note use of ag on MPI_Attr_get to
discover that a communicator has no
attribute for the keyval

123
Example: Managing tags

Problem: A library contains many objects

that need to communicate in ways that are
not known until runtime.
Messages between objects are kept separate
by using dierent message tags. How are
these tags chosen?
 Unsafe to use compile time values
 Must allocate tag values at runtime
Solution:
Use a private communicator and use an
attribute to keep track of available tags in
that communicator.

124
Caching tags on communicator

#include "mpi.h"

static int MPE_Tag_keyval = MPI_KEYVAL_INVALID;

/*
Private routine to delete internal storage when a
communicator is freed.
*/
int MPE_DelTag( comm, keyval, attr_val, extra_state )
MPI_Comm *comm;
int *keyval;
void *attr_val, *extra_state;
{
free( attr_val );
return MPI_SUCCESS;
}

125
Caching tags on communicator II

/*@
MPE_GetTags - Returns tags that can be used in
communication with a
communicator

Input Parameters:
. comm_in - Input communicator
. ntags - Number of tags

Output Parameters:
. comm_out - Output communicator. May be 'comm_in'.
. first_tag - First tag available
@*/
int MPE_GetTags( comm_in, ntags, comm_out, first_tag )
MPI_Comm comm_in, *comm_out;
int ntags, *first_tag;
{
int mpe_errno = MPI_SUCCESS;
int tagval, *tagvalp, *maxval, flag;

if (MPE_Tag_keyval == MPI_KEYVAL_INVALID) {
MPI_Keyval_create( MPI_NULL_COPY_FN, MPE_DelTag,
&MPE_Tag_keyval, (void *)0 );
}

126
Caching tags on communicator III

if (mpe_errno = MPI_Attr_get( comm_in, MPE_Tag_keyval,

&tagvalp, &flag ))
return mpe_errno;

if (!flag) {
/* This communicator is not yet known to this system,
so we
dup it and setup the first value */
MPI_Comm_dup( comm_in, comm_out );
comm_in = *comm_out;
MPI_Attr_get( MPI_COMM_WORLD, MPI_TAG_UB, &maxval,
&flag );
tagvalp = (int *)malloc( 2 * sizeof(int) );
printf( "Mallocing address %x\n", tagvalp );
if (!tagvalp) return MPI_ERR_EXHAUSTED;
tagvalp = *maxval;
MPI_Attr_put( comm_in, MPE_Tag_keyval, tagvalp );
return MPI_SUCCESS;
}

127
Caching tags on communicator IV

*comm_out = comm_in;
if (*tagvalp < ntags) {
/* Error, out of tags. Another solution would be to do
an MPI_Comm_dup. */
return MPI_ERR_INTERN;
}
*first_tag = *tagvalp - ntags;
*tagvalp = *first_tag;

return MPI_SUCCESS;
}

128
Caching tags on communicator V
/*@
MPE_ReturnTags - Returns tags allocated with MPE_GetTags.

Input Parameters:
. comm - Communicator to return tags to
. first_tag - First of the tags to return
. ntags - Number of tags to return.
@*/
int MPE_ReturnTags( comm, first_tag, ntags )
MPI_Comm comm;
int first_tag, ntags;
{
int *tagvalp, flag, mpe_errno;

if (mpe_errno = MPI_Attr_get( comm, MPE_Tag_keyval,

&tagvalp, &flag ))
return mpe_errno;

if (!flag) {
/* Error, attribute does not exist in this communicator
*/
return MPI_ERR_OTHER;
}
if (*tagvalp == first_tag)
*tagvalp = first_tag + ntags;

return MPI_SUCCESS;
}

129
Caching tags on communicator VI

/*@
MPE_TagsEnd - Returns the private keyval.
@*/
int MPE_TagsEnd()
{
MPI_Keyval_free( &MPE_Tag_keyval );
MPE_Tag_keyval = MPI_KEYVAL_INVALID;
}

130
Commentary

 Use MPI_KEYVAL_INVALID to detect when

keyval must be created
 Use flag return from MPI_ATTR_GET to
detect when a communicator needs to be
initialized

131
Exercise - Writing libraries
Objective: Use private communicators and attributes
Write a routine to circulate data to the next process,
using a nonblocking send and receive operation.
void Init_pipe( comm )
void ISend_pipe( comm, bufin, len, datatype, bufout )
void Wait_pipe( comm )

A typical use is
Init_pipe( MPI_COMM_WORLD )
for (i=0; i<n; i++) {
ISend_pipe( comm, bufin, len, datatype, bufout );
Do_Work( bufin, len );
Wait_pipe( comm );
t = bufin; bufin = bufout; bufout = t;
}

What happens if Do_Work calls MPI routines?

 What do you need to do to clean up Init pipe?
 How can you use a user-dened topology to
determine the next process? (Hint: see MPI Topo test
and MPI Cartdim get.)

132
MPI Objects

 MPI has a variety of objects

(communicators, groups, datatypes, etc.)
that can be created and destroyed. This
section discusses the types of these data and
how MPI manages them.
 This entire chapter may be skipped by
beginners.

133
The MPI Objects

MPI Request Handle for nonblocking

communication, normally freed by MPI in
a test or wait
MPI Datatype MPI datatype. Free with
MPI_Type_free.

MPI Op User-dened operation. Free with

MPI_Op_free.

MPI Comm Communicator. Free with

MPI_Comm_free.

MPI Group Group of processes. Free with

MPI_Group_free.

MPI Errhandler MPI errorhandler. Free with

MPI_Errhandler_free.

134
When should objects be freed?

Consider this code

MPI_Type_vector( ly, 1, nx, MPI_DOUBLE, &newx1 );
MPI_Type_hvector( lz, 1, nx*ny*sizeof(double), newx1,
&newx );
MPI_Type_commit( &newx );

(This creates a datatype for one face of a 3-D

decomposition.) When should newx1 be freed?

135
Reference counting

MPI keeps track of the use of an MPI object, and

only truely destroys it when no-one is using it. newx1
is being used by the user (the MPI_Type_vector that
created it) and by the MPI_Datatype newx that uses it.
If newx1 is not needed after newx is dened, it should
be freed:
MPI_Type_vector( ly, 1, nx, MPI_DOUBLE, &newx1 );
MPI_Type_hvector( lz, 1, nx*ny*sizeof(double), newx1,
&newx );
MPI_Type_free( &newx1 );
MPI_Type_commit( &newx );

136
Why reference counts

Why not just free the object?

Consider this library routine:
void MakeDatatype( nx, ny, ly, lz, MPI_Datatype *new )
{
MPI_Datatype newx1;
MPI_Type_vector( ly, 1, nx, MPI_DOUBLE, &newx1 );
MPI_Type_hvector( lz, 1, nx*ny*sizeof(double), newx1,
new );
MPI_Type_free( &newx1 );
MPI_Type_commit( new );
}

Without the MPI_Type_free( &newx1 ), it would be very

awkward to later free newx1 when new was freed.

137
Tools for evaluating programs

MPI provides some tools for evaluating the

performance of parallel programs.
These are
 Timer
 Proling interface

138
The MPI Timer

The elapsed (wall-clock) time between two

points in an MPI program can be computed
using MPI_Wtime:
double t1, t2;
t1 = MPI_Wtime();
...
t2 = MPI_Wtime();
printf( "Elapsed time is %f\n", t2 - t1 );

The value returned by a single call to

MPI_Wtime has little value.

 The times are local; the attribute

MPI WTIME IS GLOBAL may be used to determine
if the times are also synchronized with each
other for all processes in MPI COMM WORLD.

139
Proling

 All routines have two entry points: MPI ... and

PMPI ....

 This makes it easy to provide a single level of

low-overhead routines to intercept MPI calls
without any source code modications.
 Used to provide \automatic" generation of trace
les.
MPI_Send MPI_Send MPI_Send
PMPI_Send PMPI_Send

MPI_Bcast MPI_Bcast

User Program Profile Library MPI Library

static int nsend = 0;

int MPI_Send( start, count, datatype, dest, tag, comm )
{
nsend++;
return PMPI_Send( start, count, datatype, dest, tag, comm )
}

140
Writing proling routines
The MPICH implementation contains a program for
writing wrappers.
This description will write out each MPI routine that
is called.:
#ifdef MPI_BUILD_PROFILING
#undef MPI_BUILD_PROFILING
#endif
#include <stdio.h>
#include "mpi.h"

{{fnall fn_name}}
{{vardecl int llrank}}
PMPI_Comm_rank( MPI_COMM_WORLD, &llrank );
printf( "[%d] Starting {{fn_name}}...\n",
llrank ); fflush( stdout );
{{callfn}}
printf( "[%d] Ending {{fn_name}}\n", llrank );
fflush( stdout );
{{endfnall}}

The command
wrappergen -w trace.w -o trace.c
converts this to a C program. The complie the le
`trace.c' and insert the resulting object le into your
link line:
cc -o a.out a.o ... trace.o -lpmpi -lmpi

141
Another proling example
This version counts all calls and the number of bytes sent with
MPI_Send, MPI_Bsend, or MPI_Isend.
#include "mpi.h"

{{foreachfn fn_name MPI_Send MPI_Bsend MPI_Isend}}

static long {{fn_name}}_nbytes_{{fileno}};{{endforeachfn}}

{{forallfn fn_name MPI_Init MPI_Finalize MPI_Wtime}}int

{{fn_name}}_ncalls_{{fileno}};
{{endforallfn}}

{{fnall this_fn_name MPI_Finalize}}

printf( "{{this_fn_name}} is being called.\n" );

{{callfn}}

{{this_fn_name}}_ncalls_{{fileno}}++;
{{endfnall}}

{{fn fn_name MPI_Send MPI_Bsend MPI_Isend}}

{{vardecl int typesize}}

{{callfn}}

MPI_Type_size( {{datatype}}, (MPI_Aint *)&{{typesize}} );

{{fn_name}}_nbytes_{{fileno}}+={{typesize}}*{{count}}
{{fn_name}}_ncalls_{{fileno}}++;

{{endfn}}

142
Another proling example (con't)

{{fn fn_name MPI_Finalize}}

{{forallfn dis_fn}}
if ({{dis_fn}}_ncalls_{{fileno}}) {
printf( "{{dis_fn}}: %d calls\n",
{{dis_fn}}_ncalls_{{fileno}} );
}
{{endforallfn}}
if (MPI_Send_ncalls_{{fileno}}) {
printf( "%d bytes sent in %d calls with MPI_Send\n",
MPI_Send_nbytes_{{fileno}},
MPI_Send_ncalls_{{fileno}} );
}
{{callfn}}
{{endfn}}

143
Generating and viewing log les

Log les that contain a history of a

parallel computation can be very valuable
in understanding a parallel program. The
upshot and nupshot programs, provided in
the MPICH and MPI-F implementations,
may be used to view log les

144
Generating a log le

This is very easy with the MPICH

implementation of MPI. Simply replace -lmpi
with -llmpi -lpmpi -lm in the link line for
your program, and relink your program. You
do not need to recompile.
On some systems, you can get a real-time
animation by using the libraries -lampi -lmpe
-lm -lX11 -lpmpi.

Alternately, you can use the -mpilog or

-mpianim options to the mpicc or mpif77
commands.

145
Connecting several programs together

MPI provides support for connection separate

message-passing programs together through
the use of intercommunicators.

146
Sending messages between dierent programs

Programs share MPI_COMM_WORLD.

Programs have separate and disjoint
communicators.
App1 App2

Comm1 Comm2

Comm_intercomm

MPI_COMM_WORLD

147
Exchanging data between programs

 Form intercommunicator
(MPI_INTERCOMM_CREATE)
 Send data
MPI_Send( ..., 0, intercomm )

MPI_Recv( buf, ..., 0, intercomm );

MPI_Bcast( buf, ..., localcomm );

More complex point-to-point operations

can also be used

148
Collective operations

Use MPI_INTERCOMM_MERGE to create an

intercommunicator.

149
Final Comments

Additional features of MPI not covered in

this tutorial
 Persistent Communication
 Error handling

150
Sharable MPI Resources
 The Standard itself:
{ As a Technical report: U. of Tennessee.
report
{ As postscript for ftp: at info.mcs.anl.gov in
pub/mpi/mpi-report.ps.

{ As hypertext on the World Wide Web:

http://www.mcs.anl.gov/mpi

{ As a journal article: in the Fall issue of the

Journal of Supercomputing Applications
 MPI Forum discussions
{ The MPI Forum email discussions and both
current and earlier versions of the Standard
are available from netlib.
 Books:
{ Using MPI: Portable Parallel Programming
with the Message-Passing Interface, by
Gropp, Lusk, and Skjellum, MIT Press, 1994
{ MPI Annotated Reference Manual, by Otto,
et al., in preparation.

151
Sharable MPI Resources, continued
 Newsgroup:
{ comp.parallel.mpi

 Mailing lists:
{ [email protected] : the MPI Forum
discussion list.
{ [email protected] : the implementors'
discussion list.
 Implementations available by ftp:
{ MPICH is available by anonymous ftp from
info.mcs.anl.gov in the directory
pub/mpi/mpich, le mpich.tar.Z.

{ LAM is available by anonymous ftp from

tbag.osc.edu in the directory pub/lam.

{ The CHIMP version of MPI is available by

anonymous ftp from ftp.epcc.ed.ac.uk in the
directory pub/chimp/release.
 Test code repository:
{ ftp://info.mcs.anl.gov/pub/mpi/mpi-test

152
MPI-2
 The MPI Forum (with old and new participants)
has begun a follow-on series of meetings.
 Goals
{ clarify existing draft
{ provide features users have requested
{ make extensions, not changes
 Major Topics being considered
{ dynamic process management
{ client/server
{ real-time extensions
{ \one-sided" communication (put/get, active
messages)
{ portable access to MPI system state (for
debuggers)
{ language bindings for C++ and Fortran-90
 Schedule
{ Dynamic processes, client/server by SC '95
{ MPI-2 complete by SC '96

153
Summary

 The parallel computing community has cooperated

to develop a full-featured standard message-passing
library interface.
 Implementations abound
 Applications beginning to be developed or ported
 MPI-2 process beginning
 Lots of MPI material available

154