Unit-Ii PPT
Unit-Ii PPT
by
2. Eras of Computing
4. Clusters Classifications
• Parallel computers are systems that connect multiple processors to coordinate their computational efforts.
• These systems allow for the sharing of computational tasks among multiple processors.
Pfister suggests three ways to improve performance:
• SCI is an IEEE 1596-1992 standard aimed at providing low-latency distributed shared memory across a cluster.
• It is the modern equivalent of a Processor-Memory-I/O bus and LAN combined.
• SCI is a point-to-point architecture with directory-based cache coherence.
• SCI has been favored for fast distributed shared memory support, but its scalability is constrained by the current
generation of switches and its components are relatively expensive.
5.8 Myrinet
• Distributed applications require diverse communication needs, ranging from reliable point-to-point to
unreliable multicast communications.
• The communications infrastructure supports protocols for bulk data transport, streaming data, group
communications, and distributed objects.
• Communication services provide basic mechanisms for transporting administrative and user data, and provide
quality of service parameters like latency, bandwidth, reliability, fault-tolerance, and jitter control.
Network services are typically designed as a hierarchical stack of protocols, with each protocol layer exploiting
the services provided by the protocols below it.
• Traditionally, operating system services were used for communication between processes in message passing
systems, involving expensive operations.
• Clusters with special networks/switch like Myrinet use lightweight communication protocols for fast
communication among nodes, bypassing the operating system and providing direct, user-level access to the
network interface.
• Network services are often built from a low-level communication API, supporting a wide range of high-level
communication libraries and protocols.
7. Cluster Middleware and Single System Image (SSI)
• Hardware Layer
• Benefits of SSI
• Complete Transparency
• Scalable Performance
• Enhanced Availability
8. Key Services of SSI and Availability Infrastructure
• Services provided by RMS include Process Migration, Checkpointing, Scavenging Idle Cycles, Fault
Tolerance, Minimization of Impact on Users, Load Balancing, and Multiple Application Queues.
• There are many commercial and research packages available for RMS, with several in-depth reviews of the
available RMS systems.
Project Commercial Systems URL
LSF http://www.platform.com/
CODINE http://www.genias.de/products/codine/tech_desc.html
NQE http://www.cray.com/products/software/nge/
CONDOR http://www.cs.wisc.edu/condor/
GNQS http://www.gnqs.org/
DQS http://www.scri.fsu.edu/~pasko/dqs.html
PRM http://gost.isi.edu/gost-group/products/prm/
PBS http://pbs.mrj.com/
10. Programming Environments and Tools
10.1 Threads
• Threads are a popular paradigm for concurrent programming on both uniprocessor and multiprocessor
machines.
• They exploit the asynchronous behavior of an application for overlapping computation and communication.
• Threads are potentially portable, with an IEEE standard for POSIX threads interface, pthreads.
• Programming languages like Java have built-in multithreading support, enabling easy development of
multithreaded applications.
10.2 Message Passing Systems (MPI and PVM)
• Message passing libraries allow efficient parallel programs to be written for distributed memory systems.
• PVM is both an environment and a message passing library, used to run parallel applications on systems
ranging from high-end supercomputers to clusters of workstations.
• MPI is a message passing specification designed to be standard for distributed memory parallel computing
using explicit message passing.
• MPI is available on most of the HPC systems, including SMP machines.
10.3 Distributed Shared Memory (DSM) Systems
• Message passing is the most efficient programming paradigm on distributed memory systems.
• DSM enables shared-variable programming and can be implemented using software or hardware solutions.
• Software DSM systems are usually built as a separate layer on top of the communications interface, while hardware DSM systems have better
performance, no burden on user and software layers, fine granularity of sharing, extensions of the cache coherence schemes, and increased
hardware complexity.
• Parallel debuggers and profilers are essential for efficient development of high performance applications.
• Most vendors of HPC systems provide debuggers and performance analyzers for their platforms.
• These tools should work in a heterogeneous environment, enabling parallel application development on a NOW and production runs on a
• Threads
• Debuggers
• TotalView
• GCAs are fundamental problems in science and engineering with broad economic and scientific impact.
• Examples include massive crystallographic and micro tomographic structural problems, protein dynamics,
bio catalysis, relativistic quantum chemistry, virtual materials design, global climate modeling, and discrete
event simulation.
Concepts covered in Unit- III
First Part
1. Introduction(Computing Power Limitations and Parallel Computing)
2. Eras of Computing
4. Clusters Classifications