Pattern Programming Barry Wilkinson University of North Carolina

Pattern Programming Barry Wilkinson University of North Carolina Charlotte CCI Friday Seminar Series April 13 th, 2012

Acknowledgment This work was initiated by Jeremy Villalobos and described in his Ph. D thesis: “RUNNING PARALLEL APPLICATIONS ON A HETEROGENEOUS ENVIRONMENT WITH ACCESSIBLE DEVELOPMENT PRACTICES AND AUTOMATIC SCALABILITY, ” UNC-Charlotte, 2011. 2

Problem Addressed • To make parallel programming more useable and scalable. • Parallel programming is writing programs for solving problems using multiple computers, processors and cores, including with physically distributed computers. • A very long history but still a challenge. • Traditional approaches involve explicitly specifying message-passing with low-level tools such as MPI and thread parallelism with Open. MP and Open. CL. • Still not mainstream as it should be with the introduction of multicore processors 3

Pattern Programming Concept Programmer constructs his application using established computational or algorithmic “patterns” that provide a structure. Patterns are widely applicable. What patterns are we talking about? • Low-level algorithmic patterns that might be embedded into a program such as fork-join, broadcast/scatter/gather. • Higher level algorithm patterns forming a complete program such as workpool, pipeline, stencil, map-reduce. We concentrate upon higher-level “computational/algorithm ” level patterns rather than lower level patterns. 4

Some Patterns Workpool Workers Master Two-way connection Compute node Source/sink Derived from Jeremy Villalobos’s Ph. D thesis defense 5

Pipeline Stage 1 Stage 2 Stage 3 Workers Master One-way connection Two-way connection Compute node Source/sink 6

Divide and Conquer Divide Merge Two-way connection Compute node Source/sink 7

Stencil Synchronous Two-way connection Compute node Source/sink 8

All-to-All Two-way connection Compute node Source/sink 9

Note on Terminology “Skeletons” Sometimes term “skeleton” used to describe “patterns”, especially directed acyclic graphs with a source, a computation, and a sink. We do not make that distinction and use the term “pattern” whether directed or undirected and whether acyclic or cyclic. This is done elsewhere. 10

Skeletons/Patterns • Advantages • Implicit parallelization • Avoid deadlocks • Avoid race conditions • Reduction in source code size (lines of code) • Abstracts the Grid/Cloud environment • Disadvantages • Takes away some of the freedom from the user programmer • New approach to learn • Performance reduced (5% on top of MPI) Derived from Jeremy Villalobos’s Ph. D thesis defense 11

More Advantages/Notes • “Design patterns” have been part of software engineering for many years. . – – Reusable solutions to commonly occurring problems * Patterns provide guide to “best practices”, not a final implementation Provides good scalable design structure to parallel programs Can reason more easier about programs • Hierarchical designs with patterns embedded into patterns, and pattern operators to combine patterns. • Leads to an automated conversion into parallel programs without need to write with low level message-passing routines such as MPI. * http: //en. wikipedia. org/wiki/Design_pattern_(computer_science) 12

Previous/Existing Work • Patterns/skeletons explored in several projects. • Universities: – University of Illinois at Urbana-Champaign and University of California, Berkeley – University of Torino/Università di Pisa Italy –. . . • Industrial efforts – Intel – Microsoft –… 13

Universal Parallel Computing Research Centers (UPCRC) University of Illinois at Urbana-Champaign and University of California, Berkeley with Microsoft and Intel in 2008 (with combined funding of at least $35 million). – Co-developed OPL (Our Pattern Language), a pattern language for parallel programming – Promoted patterns in Workshop on Parallel Programming Patterns, Para. PLo. P 2009, 2010, and 2011, and Pattern Languages for Programs conference Plo. P’ 2009 14

UPCRC Patterns Group of twelve computational patterns identified: • • • Finite State Machines Circuits Graph Algorithms Structured Grid Dense Matrix Sparse Matrix • • • Spectral (FFT) Dynamic Programming Particle Methods Backtrack Graphical Models Unstructured Grid in seven general application areas 15

Closest to our work http: //calvad os. di. unipi. it/d okuwiki/doku. php? id=ffnam espace: about University of Torino, Italy /Università di Pisa 16

Intel Focused on very low level patterns such as fork-join, and provides constructs for them in: • Intel Threading Building Blocks (TBB) – Template library for C++ to support parallelism • Intel Cilk plus – Compiler extensions for C/C++ to support parallelism • Intel Array Building Blocks (Ar. BB) – Pure C++ library-based solution for vector parallelism Above are somewhat competing tools obtained through takeovers of small companies. Each implemented differently. 17

New book due out 2012 from Intel authors “Structured Parallel Programming: Patterns for Efficient Computation, ” Michael Mc. Cool, James Reinders, Arch Robison, Morgan Kaufmann, 2012 Focuses on Intel tools B. Wilkinson was a reviewer for this book. 15. 18

Using patterns with Microsoft C# http: //www. microsoft. com/download/en/det ails. aspx? displaylang=en&id=19222 Again very low-level with patterns such as parallel for loops. 19

Our approach (Jeremy Villalobos’ UNC-C Ph. D thesis) Focuses on a few patterns of wide applicability (workpool, pipelined, stencil, and dense matrix patterns) but Jeremy took it much further than UPCRC and Intel. He developed a higher-level framework called “Seeds” Uses pattern approach to automatically distribute code across geographical sites and execute the parallel code. 20

“Seeds” Parallel Grid Application Framework Some Key Features • Pattern-programming (Java) user interface • Self-deploys on computers, clusters, and geographically distributed computers • Load balances • Three levels of user interface http: //coit-grid 01. uncc. edu/seeds/21

Seeds Development Layers Basic Advanced: Used to add or extend functionality such as: Intended for programmers that have basic parallel computing background Based on skeletons and patterns Create new patterns Optimize existing patterns or Adapt existing pattern to non-functional requirements specific to the application Expert: Used to provide basic services: Deployment Security Communication/Connectivity Changes in the environment Derived from Jeremy Villalobos’s Ph. D thesis defense 22

Deployment • Deployment with Globus – Globus GSIFTP to transfer “seeds” folder – GRAM to submit job to run seed nodes • Deployment with SSH – now preferred – Globus will be depreciated in Seeds 23

Basic User Programmer Interface To create and execute parallel programs, programmer selects a pattern and implements three principal Java methods: • Diffuse method – to distribute pieces of data. • Compute method – the actual computation Diffuse Compute Gather • Gather method – used to gather the results Bootstrap class Programmer also has to fill in details in a “bootstrap” class to deploy and start the framework. The framework self-deploys on a geographically distributed platform and executes pattern. 24

Example: Deploy a workpool pattern to compute p using Monte Carlo method Monte Carlo p calculation • Basis on Monte Carlo calculations is use of random selections • In this case, circle formed with a square • Points within square chosen randomly • Fraction of points within circle = p/4 • Only one quadrant used in code 25

Complete code for computation package edu. uncc. grid. example. workpool; import java. util. Random; import java. util. logging. Level; import edu. uncc. grid. pgaf. datamodules. Data. Map; import edu. uncc. grid. pgaf. interfaces. basic. Workpool; import edu. uncc. grid. pgaf. p 2 p. Node; public class Monte. Carlo. Pi. Module extends Workpool { private static final long serial. Version. UID = 1 L; private static final int Double. Data. Size = 1000; double total; int random_samples; Random R; public Monte. Carlo. Pi. Module() { R = new Random(); } @Override public void initialize. Module(String[] args) { total = 0; Node. get. Log(). set. Level(Level. WARNING); // reduce verbosity for logging random_samples = 3000; // set number of random samples } public Data Compute (Data data) { // input gets the data produced by Diffuse. Data() Data. Map<String, Object> input = (Data. Map<String, Object>)data; // output will emit the partial answers done by this method Data. Map<String, Object> output = new Data. Map<String, Object>(); Long seed = (Long) input. get("seed"); // get random seed Random r = new Random(); r. set. Seed(seed); Long inside = 0 L; for (int i = 0; i < Double. Data. Size ; i++) { double x = r. next. Double(); double y = r. next. Double(); double dist = x * x + y * y; if (dist <= 1. 0) { ++inside; } } output. put("inside", inside); // store partial answer to return to Gather. Data() return output; } public Data Diffuse. Data (int segment) { Data. Map<String, Object> d =new Data. Map<String, Object>(); d. put("seed", R. next. Long()); return d; // returns a random seed for each job unit } public void Gather. Data (int segment, Data dat) { Data. Map<String, Object> out = (Data. Map<String, Object>) dat; Long inside = (Long) out. get("inside"); total += inside; // aggregate answer from all the worker nodes. } public double get. Pi() { // returns value of pi based on the job done by all the workers double pi = (total / (random_samples * Double. Data. Size)) * 4; return pi; } public int get. Data. Count() { return random_samples; Note: No message passing (MPI etc) 26

package edu. uncc. grid. example. workpool; import java. io. IOException; import net. jxta. pipe. Pipe. ID; import edu. uncc. grid. pgaf. Anchor; import edu. uncc. grid. pgaf. Operand; import edu. uncc. grid. pgaf. Seeds; import edu. uncc. grid. pgaf. p 2 p. Types; public class Run. Monte. Carlo. Pi. Module { public static void main(String[] args) { try { Monte. Carlo. Pi. Module pi = new Monte. Carlo. Pi. Module(); Seeds. start( "/path/to/seeds/seed/folder" , false); Pipe. ID id = Seeds. start. Pattern(new Operand( (String[])null, new Anchor( "hostname" , Types. Data. Flow. Roll. SINK_SOURCE), pi ) ); System. out. println(id. to. String() ); Seeds. wait. On. Pattern(id); System. out. println( "The result is: " + pi. get. Pi() ) ; Seeds. stop(); } catch (Security. Exception e) { e. print. Stack. Trace(); } catch (IOException e) { e. print. Stack. Trace(); } catch (Exception e) { e. print. Stack. Trace(); } } Bootstrap class This code deploys framework and starts execution of pattern Different patterns have similar code 27

Compiling/executing • Can be done on the command line (ant script provided) or through an IDE (Eclipse) 28

Another example Bubble sort with the pipeline pattern From Jeremy Villalobos’ Ph. D thesis. See thesis for more details and results Static test - Program executed using a fixed number of cores as given. Dynamic test - Number of cores dynamically changed during execution to improve performance. Platform: Dell 900 server with four quad-core processors and 64 GB shared memory. 29

Pattern operators Can combine patterns. Example: Adding Stencil and All-to-All synchronous pattern Example use: Heat distribution simulation (Laplace’s eq. ) • Multiple cells on a stencil pattern work in a loop parallel fashion, computing and synchronizing on each iteration. • However, every x iterations, they must implement an all-to-all communication pattern to run an algorithm to detect termination. Directly from Jeremy Villalobos’s Ph. D thesis 30

Tutorial page 15. 31

15. 32

Download page 33

Open Source Apache License in progress http: //code. google. com/p/seeds-parallel-pattern-framework/ 34

Work in progress • Tutorial on using all-to-all pattern to solve the n-body problem, and other documentation • Documentation on advanced layer to enable programmers to develop their own patterns • Exploring various ways to enhance and expand the work, using for example Aparapi http: //code. google. com/p/aparapi/ 35

ITCS 4145/5145 Parallel Programming • Pattern programming to be introduced into ITCS 4145/5145 Parallel Programming in Fall 2012. To be taught on NCREN in same way as ITCS 4/5146 Grid Computing with instructors at UNCCharlotte and UNCWilmington but in Fall 2012 just two sites. Subsequently will be offered across NC. Regional/national workshops planned External funding for this work pending. 36

Pattern Programming Research Group • 2011 – Jeremy Villalobos (Ph. D awarded, continuing involvement) – Saurav Bhattara (MS thesis, graduated) • Spring 2012 – Yawo Adibolo (ITCS 6880 Individual Study) – Ayay Ramesh (ITCS 6880 Individual Study) • Fall 2012 – Haoqi Zhao (MS thesis) Openings! Loosely related Spring 2012: Tim Lukacik and Phil Chung (UG senior projects on Eclipse PTP) 37

Some publications • Jeremy F. Villalobos and Barry Wilkinson, “Skeleton/Pattern Programming with an Adder Operator for Grid and Cloud Platforms, ” The 2010 International Conference on Grid Computing and Applications (GCA’ 10), July 12 -15, 2010, Las Vegas, Nevada, USA. • Jeremy F. Villalobos and Barry Wilkinson, “Using Hierarchical Dependency Data Flows to Enable Dynamic Scalability on Parallel Patterns, ” High-Performance Grid and Cloud Computing Workshop, 25 th IEEE International Parallel & Distributed Processing Symposium, Anchorage (Alaska) USA, May 16 -20, 2011. I regard this conference as a top conference in the field Also presented by B. Wilkinson as Session 4 in “Short Course on Grid Computing” Jornadas Chilenas de Computación, INFONOR-CHILE 2010, Nov. 18 th - 19 th, 2010, Antofagasta, Chile. http: //coitweb. uncc. edu/~abw/Infornor-Chile 2010/Grid. Workshop/index. html 38

Questions 39