1 UNC-Charlotte’s Grid Computing “Seeds” framework 1 © 2011 Jeremy Villalobos /B. Wilkinson Fall 2011 Grid computing course. Slides10-1.ppt Modification date: Nov 16, 2011 Acknowledgment The “Seeds” framework was developed by: Jeremy Villalobos, BS, MS, PhD (UNC-Charlotte) as part of his PhD work at UNC-Charlotte. The following slides based upon materials provided by Jeremy Villalobos.

Challenges in Running Parallel Applications on The Grid Heterogeneity Connectivity Security Programmability Performance and Scalability 2 The Seeds framework particularly addresses programmability and scalability

MPI and OpenMP OpenMP (Shared Memory) Easy to program High performance Does not scale MPI (Distributed Memory) Hard to program High performance Can scale 3

“Seeds Distributed Computing Framework Developed at UNC- Charlotte by Jeremy Villalobos Raises level of programming for ease of programmability based upon pattern programming. Programmer first identifies an appropriate parallel pattern or patterns to solve the problem - the patterns in this context being for example workpool, pipeline, synchronous, mesh/stencil, …. 4 Workpool pattern

“Seeds Distributed Computing Framework Programmer then simply implements certain Java interface methods, principally:  data diffuse  computation, and  data gather and framework automatically distributes tasks across distributed computers and processor cores. Key aspects  Programmer does not write programs at low level of message passing MPI or thread-based APIs (not initially anyway).  Patterns implemented automatically and code is executed on a distributed computing platform. 5

“Skeletons” and “Patterns” Note. In the literature, term “skeleton” sometimes used to describe “patterns”, especially directed acyclic graphs with a source, a computation, and a sink. 6

Skeletons Skeletons are: Functional programming constructs, Stateless Data-parallel Resemble trees 7

Pattens Patterns are: State-full, Synchronous loop-parallel Also data-parallel 8

Skeletons/Patterns A dvantages Implicit parallelization Avoid deadlocks Avoid race conditions Reduction in code size [3] Abstracts the Grid/Cloud environment Disadvantages Takes away some of the freedom from the user programmer New approach to learn Performance told (5% for skeletons on top of MPI [PASM]) 9

Hands-on session Deploy a Seeds workpool skeleton to compute  using Monte Carlo method – all code given. Just local computer used for this session although remote computers could be used. Needs Seeds installed. For session, will also need Java, ant, and Eclipse installed First prepared ant script used to deploy workpool Then Eclipse IDE used to run same program Later Eclipse example for numerical integration - code partially given. Will need to fill in missing parts. 10

11 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 =  /4  calculation

12 Can limit calculation to one quadrant and get same result Actually computes an integral

Seed Skeletons/ Patterns 13

14 Code Initial part package edu.uncc.grid.example.workpool; import java.util.Random; import java.util.logging.Level; import edu.uncc.grid.pgaf.datamodules.Data; import edu.uncc.grid.pgaf.datamodules.DataMap; import edu.uncc.grid.pgaf.interfaces.basic.Workpool; import edu.uncc.grid.pgaf.p2p.Node; public class MonteCarloPiModule extends Workpool { private static final long serialVersionUID = 1L; private static final int DoubleDataSize = 1000; double total; int random_samples; Random R; public MonteCarloPiModule() { R = new Random(); public void initializeModule(String[] args) { total = 0; // reduce verbosity for logging information Node.getLog().setLevel(Level.WARNING); // set number of random samples random_samples = 3000; }

15 Compute method public Data Compute(Data data) { // input gets the data produced by DiffuseData() DataMap input = (DataMap )data; // output will emit the partial answers done by this method DataMap output = new DataMap (); Long seed = (Long) input.get("seed"); // get random seed Random r = new Random(); r.setSeed(seed); Long inside = 0L; for (int i = 0; i < DoubleDataSize ; i++) { double x = r.nextDouble(); double y = r.nextDouble(); double dist = x * x + y * y; if (dist <= 1.0) { ++inside;// = 1L; } output.put("inside", inside);// store partial answer to return //to GatherData() return output; }

Diffuse and gather methods public Data DiffuseData(int segment) { DataMap d =new DataMap (); d.put("seed", R.nextLong()); return d; // returns a random seed for each job unit } public void GatherData(int segment, Data dat) { DataMap out = (DataMap ) dat; Long inside = (Long) out.get("inside"); total += inside; // aggregate answer from all the worker nodes. } public double getPi() { // returns value of pi based on the job done by all the workers double pi = (total / (random_samples * DoubleDataSize)) * 4; return pi; } public int getDataCount() { return random_samples; } 16

Running example using ant script Get prepared package PiApprox.zip Contains everything needed:  Seeds  P code  Numerical integration code Run ant script Will compile all software and run PiApprox program 17

Using Eclipse Advantage of using the ant script first is everything is deployed (seed folder, etc.) 18

Additional Features of Framework 19

Seeds Specifications Topology: Network Overlay/Hybrid Network Connectivity: Direct, Through NAT, and Shared memory Programming Style: Skeleton/Pattern Based Parallel Memory Management: Message- passing with use of shared memory if available Self-deployment: Using Java Cog Kit and Globus Load-balancing: expected on the patterns 20

Seeds Development Layers Basic Intended for programmers that have basic parallel computing background Based on Skeletons and patterns Advanced: Used to add or extend functionality like: 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 21

Seeds Development Layers Basic Intended for programmers that have basic parallel computing background Based on Skeletons and patterns Advanced: Used to add or extend functionality like: 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 22

Nested Pattern Structures It may be that a single pattern will not suffice for a larger problem. Framework offers nested patterns and facilities to break a single pattern into multiple patterns. 23

Pattern Operators Also have pattern operators that can take two patterns and merge them into a single pattern. 24 Adding a Stencil (left) to an all-to-all (center) to produce hybrid pattern (right).

Some results 25 Bubble sort using a pipeline on shared memory with Seeds framework Pipeline pattern for Bubblesort, tested on the Dell 900 server with four quad-core processors and 64GB shared memory (coit-grid05.uncc.edu). Static test, - program was executed using a fixed number of cores as given. Dynamic tests -- dynamically change the number of cores during execution to improve performance.

Questions 26