Presentation is loading. Please wait.

Presentation is loading. Please wait.

A Pattern Language for Parallel Programming Beverly Sanders University of Florida.

Similar presentations


Presentation on theme: "A Pattern Language for Parallel Programming Beverly Sanders University of Florida."— Presentation transcript:

1 A Pattern Language for Parallel Programming Beverly Sanders University of Florida

2 Overview of talk History of pattern languages Motivation for a pattern language for parallel programming Pattern example: Reengineering for Parallelism Tour through the entire pattern language via a programming example

3 History ‘60s and ‘70s Berkeley architecture professor Christopher Alexander 253 patterns for city planning, landscaping, and architecture Attempted to capture principles for “living” design.

4 Example: Six Foot Balcony Balconies less than six feet deep are hardly ever used Discussion of what makes a good balcony Therefore: Whenever you build a balcony or a porch, always make it at least six feet deep. If possible, recess at least a part of it into the building so that it is not cantilevered out and separated from the building by a simple line, and enclose it partially

5

6 A new approach to design Not just a collection of patterns, but a pattern language –Patterns lead to other patterns –Patterns are hierarchical and compositional –Embodies design methodology and vocabulary Small impact on architectural practice

7 Patterns in Object-oriented Programming OOPSLA’87 Kent Beck and Ward Cunningham 1995 Design Patterns: Elements of Reusable Object-Oriented Software Gang of Four (GOF): Gamma, Helm, Johnson, Vlissides, –catalog of patterns –Creation, structural, behavioral PLoP Conferences

8 GOF Pattern Example Behavioral Pattern: Visitor –Separate the structure of an object collection from the operations performed on that collection. –Example: Abstract syntax tree in a compiler Multiple node types (declaration, command, expression, etc.) Action during traversal depends on both type of node and compiler pass (type checking, code generation) Can add new functionality by implementing new visitor without modifying AST code.

9 Impact of GOF book Good solutions to frequently recurring problems New vocabulary Pattern catalog Significant influence on object-oriented programming!

10 Design Pattern High quality solution to frequently recurring problem in some domain Each pattern has a name, providing a vocabulary for discussing the solutions Written in prescribed format to allow the reader to quickly understand the solution and its context

11 A pattern format Name Also known as ProblemContextForcesSolution Examples and known uses Related patterns …

12 Pattern Language Carefully structured collection of patterns Structure embodies a design methodology and leads user through the language so that complex designs can be developed using patterns Provides domain specific advice to the designer Not a programming language

13 Parallel Programming Parallel hardware becoming increasingly mainstream and inexpensive –Multicore CPUs in desktop PCs and servers –Clusters Software to fully exploit the hardware currently rare (except specialized area of high performance computing) Can a pattern language providing guidance for the entire development process make parallel programming easier?

14 Structure of the pattern language –Reengineering for Parallelism pattern for dealing with legacy sequential code 4 Design spaces –Finding Concurrency Help designer expose exploitable concurrency— find high level task and data decomposition –Algorithm Structure Help designer map tasks to processes or threads to best take advantage of the potential concurrency

15 Structure of the pattern language, continued –Supporting Structures Code structuring patterns Distributed and thread-safe data structures –Implementation Mechanisms Low level mechanisms used to write parallel programs 3 categories of mechanisms –UE (process/thread) Management –Synchronization –Communication

16 Starting with legacy sequential application? Reengineering for Parallelism pattern provides guidance to –Manage the process –Determine what to change We’ll look at this as an example pattern

17 Reengineering for Parallelism Problem: How can existing applications be parallelized using PLPP to improve performance by making use of parallel hardware? Context: We have legacy code that cannot be rewritten from scratch, need to improve performance… Forces: –User base has expectations for behavior –Existing application may not be fully understood –Amdahl’s law pushes programmer to avoid sequential bottlenecks at any cost, which may imply wholesale restructuring of the program –Starting point is working code that embodies significant programming work, bug fixes, and knowledge. Minimizing changes is desirable. It is rarely feasible to make sweeping rewrites. –Concurrency introduces new classes of errors that are hard to detect and make software difficult to validate.

18 Solution:Preparation Survey the landscape –Pattern provides a list of questions to help assess existing code –Many are the same as in any reengineering project –Is program numerically well-behaved? Define the scope and get users’ buy-in –Required precision of results –Input range –Performance –Feasibility (back of envelope calculations) Define a testing protocol

19 Solution: Continued Identify hot spots—where is most of the time spent? –Look at code –Use profiling tools Parallelization –Start with hot spots first –As much as possible, make sequence of small changes, each followed by testing –Use PLPP patterns (pattern provides guidance)

20 Reengineering for Parallelism Pattern, continued Extended example Discussion of related patterns –Patterns for legacy code –Patterns for parallel programming

21 Example: Molecular dynamics Simulate motion in large molecular system Example application: how protein interacts with drug Forces –Bonded forces within a molecule –Long-range forces between molecules Not tractable N 2 Use cutoff method—only consider forces from neighbors that are “close enough”

22 Sequential Molecular dynamics simulation real atoms(3,N) real atoms(3,N) real force(3,N) real force(3,N) int neighbors(2,M) int neighbors(2,M) loop over time steps Compute bonded forces Compute neighbors Compute long-range forces Update position … end loop

23 Starting with legacy sequential code? If so start with the Reengineering for Parallelism pattern Next: Finding Concurrency Design Space

24 Decomposition Patterns Finding Concurrency Design Space Dependency Analysis Patterns Design Evaluation

25 Decomposition Patterns Finding Concurrency Design Space Dependency Analysis Patterns Design Evaluation Task DecompositionData Decomposition

26 Molecular dynamics decomposition Each function is a loop over atoms Suggests task decomposition with each task corresponding to a loop iteration (update of an atom) –tasks for bonded forces –tasks for long -range forces –tasks to update positions –tasks to compute neighbor list Data shared between the tasks

27 Decomposition Patterns Finding Concurrency Design Space Dependency Analysis Patterns Design Evaluation Group tasksOrder tasksData Sharing

28 Molecular dynamics dependency analysis Bonded forces Neighbor list Update position Long-range forces next time step

29 Molecular dynamics dependency analysis Bonded forces Neighbor list Update position Long-range forces next time step atoms(3,N) forces(3,N) neighbors Read Write Accumulate

30 Decomposition Patterns Finding Concurrency Design Space Dependency Analysis Patterns Design Evaluation Suitability for target platform Design Quality (flexibility, efficiency, simplicity) Preparation for the next phase

31 Design evaluation for molecular dynamics Target architecture for example: distributed memory cluster, message passing Data sharing has enough special properties (read only, accumulate, temporal constraints) that we should be able to make it work in a distributed memory environment. Design seems OK, move to next design space

32 Algorithm structure design space Map tasks to Units of Execution (threads or processes) Target platform properties –number of UEs –communication between UEs Major organizing principle

33 Organize by tasks? Recursive? Task Parallelism Divide and Conquer yes no

34 Organize by data? Recursive? Geometric Decomposition Recursive Data yes no

35 Organize by ordering? Regular? Event-based Coordination Pipeline yes no

36 Algorithm structure for molecular dynamics Organized by task Task decomposition pattern –Granularity: decide bonded forces not worth parallelizing now. –Load balancing: static OK, partition iterations of original loop (over atoms) to UEs –Termination: easy since number of UEs can be determined in advance

37 Separable dependencies Multiple tasks update force array concurrently by adding to its value. This type of update is called accumulation Allows dependencies to be separated from concurrent part of computation –Each UE gets a local copy of data –Updates local copy –After local updates completed, reduce (combine results using associative operator)

38 Supporting Structures Design Space A intermediate stage between algorithm structures and implementation mechanisms (Similar level to GOF) Program structuring patterns –SPMD, Fork/Join, Loop Parallelism, Master/Worker Data structures –Shared queue, Distributed array, Shared data

39 Choose SPMD Pattern Single program multiple data. –Each UE executes exactly the same program –Uses process ID to determine behavior –Issues: replicate or partition data, computation?

40 Replicate or partition data in MD? –Replicate atoms, force. –Partition neighbor list Duplicate non-parallelized parts of computation, or designate one process to compute? –Duplicate all computation except I/O.

41 Parallel Simulation real atoms(3,N) real atoms(3,N) real force(3,N) real force(3,N) int neighbors(2,M) int neighbors(2,M) myID = getProcessID myID = getProcessID nprocs = getNumberProcesses nprocs = getNumberProcesses loop over time steps Compute bonded forces //replicate computation Compute neighbors //only for atoms assigned to myID //only for atoms assigned to myID Compute long range forces globalSum(N,&forces) //reduction to combine all force arrays //reduction to combine all force arrays Update position … end loop if (myid == printID) printResults

42 Implementation Mechanisms Describes low level mechanisms used to write parallel programs 3 categories of mechanisms –UE (process/thread) Management –Synchronization –Communication Not in pattern format We discuss OpenMP, MPI, and Java

43 Implement Simulation For our target platform, MPI is the best choice Need to add standard code for initialization and MPI specific reduction operator to previous solution

44 Pattern languages evolve A pattern language should not be considered a static document: –Evaluate and revise –Extend with new patterns: new parallel programming models, specific application domains –We added the Reengineering for Parallelism pattern as a result of feedback from readers

45 For more information Mattson, Sanders, and Massingill. Patterns for Parallel Programming. Addison-Wesley Software Patterns Series. 2005 Reengineering for Parallelism, PLoP05 www.cise.ufl.edu/research/ParallelPatterns

46 Acknowledgements and collaborators The pattern language is joint work with: –Tim Mattson, Intel –Berna Massingill, Trinity University Supported by NSF and Intel


Download ppt "A Pattern Language for Parallel Programming Beverly Sanders University of Florida."

Similar presentations


Ads by Google