Lecture_3 - College of Information Technology Master...

Info iconThis preview shows pages 1–9. Sign up to view the full content.

View Full Document Right Arrow Icon
College of Information Technology Master Program in Scientific Computing Scientific Computing II (SCOM6301) Introduction to Parallel Computing Lecture 3 Parallel Computer Architectures CH03
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Overview The principal challenge of parallel programming is to decompose the program into subcomponents that can be run in parallel. In order to understand some of the low-level issues of decomposition, the programmer must have a simplified view of parallel machine architecture.
Background image of page 2
Decomposition Strategies Standard parallel architectures support a variety of decomposition strategies. Two main strategies Decomposition by task (task parallelism) Decomposition by data (data parallelism) We will concentrate on data parallelism as it is the most common strategy for scientific programs on parallel machines.
Background image of page 3

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Decomposition Strategies --- Cont In data parallelism, the application is decomposed by subdividing the data space over which it operates and assigning different processors to the work associated with different data subspaces. This strategy involves some data sharing at the boundaries, and the programmer is responsible for ensuring that this data sharing is handled correctly that is, data computed by one processor and used by another are correctly synchronized.
Background image of page 4
Programming Models The two most common models used by programmers are : The shared-memory model , in which it is assumed that all data structures are allocated in a common space that is accessible from every processor. The message-passing model , in which each processor (or process) is assumed to have its own private data space, and data must be explicitly moved between spaces as needed.
Background image of page 5

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
The message-passing model In the message-passing model, data structures are distributed across the processor memories; if a processor needs to use a data item that is not stored locally, the processor that owns that data item must explicitly “send” it to the requesting processor. The latter must execute an explicit “receive” operation, which is synchronized with the send, before it can use the communicated data item.
Background image of page 6
Performance Issues To achieve high performance on parallel machines, the programmer must be concerned with scalability and load balance . Generally, an application is thought to be scalable if larger parallel configurations can solve proportionally larger problems in the same running time as smaller problems on smaller configurations.
Background image of page 7

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
-- Cont. Load balance typically means that the processors have roughly the same amount of work, so that no one processor holds up the entire solution. To balance the computational load on a machine with processors of equal power, the programmer must divide the work and communications evenly. This can be challenging in applications applied to
Background image of page 8
Image of page 9
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 09/24/2009 for the course CS 525 taught by Professor Rjyosy during the Spring '09 term at Central Mich..

Page1 / 34

Lecture_3 - College of Information Technology Master...

This preview shows document pages 1 - 9. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online