1 / 16

Matthias Vodel Wolfram Hardt

Matthias Vodel Wolfram Hardt vodel@cs.tu-chemnitz.de { hardt }@cs.tu-chemnitz.de. Content. Motivation MPI2Java Concept Performance Analysis SimANet Application Scenarios. APCC 2010 Matthias Vodel , University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de - 2 -.

brie
Télécharger la présentation

Matthias Vodel Wolfram Hardt

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Matthias Vodel Wolfram Hardt vodel@cs.tu-chemnitz.de{ hardt }@cs.tu-chemnitz.de Parallel High-Performance Applications with MPI2Java - A Capable Java Interface for MPI 2.0 Libraries -

  2. Content • Motivation • MPI2Java Concept • Performance Analysis • SimANet Application Scenarios APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -2-

  3. High Performance Computing Motivation• Concept • Analysis • SimANet • High complex computing challenges in the field of: • Mathematics • Physics • Chemistry • Engineering • Informatics • Weather forecast Massive parallel computing power State of the art hardware Modern implementation languages APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -3-

  4. MPI Standard Motivation• Concept • Analysis • SimANet • Message Passing Interface • Communication protocol for distributed • computing environments • MPI specifies communication features • Implemented in libraries • Usage by an defined API De facto standard for computer cluster programming MPICH represents a implementation of the MPI specification Performance-optimised libraries for C and Fortran Missing up-to-date implementation for modern languages (object-orientation, template-based design patterns) APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -4-

  5. MPI Support for Java Motivation• Concept • Analysis • SimANet + platform-independent + no C-libraries required + minimal maintenance costs Pure Java Implementations JNI Implementations + usage of related, performance-optimised libraries + less implementation work + significant faster • complete reimplementation of the entire MPI specification • performance problems (higher abstraction level, minimal optimisations) • maintenance to keep the C-libraries • up-to-date • Installation on the target system hardware is more complex (library binaries not platform-independent) APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -5-

  6. MPI Support for Java Motivation• Concept • Analysis • SimANet Pure Java Implementations JNI Implementations APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -6-

  7. Why implementing anew MPI-Java interface?? Motivation• Concept • Analysis • SimANet Since 2003: no further development of existing projects JMPI significant slower than mpiJava mpiJava – bad exception / error handling – missing 64bit support  actual cluster computer environment not supported! All Java-MPI libraries are not able to use MPI-2 features APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -7-

  8. MPI2Java - Structure Motivation• Concept • Analysis • SimANet APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -8-

  9. MPI2Java - Interface Motivation• Concept • Analysis • SimANet = in Java implementiert = in C implementiert MPI.java libMPI.c public class MPI { public static String[] Init(...) { ... nativeInit(...); ... } } #include <mpi.h> jobjectArraynativeInit(...) { ... MPI_Init(...); ... } MpiTest.java ... MPI.Init(...); ... MPI.COMM_WORLD.Send(...); ... Comm.java libComm.c public class Comm { public Status Send(...) { ... nativeSend(...); ... } } #include <mpi.h> jobject nativeSend(...) { ... MPI_Send(...); ... } APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -9-

  10. PC Workstation Environment Motivation• Concept • Analysis • SimANet PC workstation - 3Ghz Pentium 4, 2GB RAM, 100Mbit/s fast Ethernet - Scientific Linux 4.4, 2.6.18 Kernel Throughput measurements Pingpong measurements APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -10-

  11. CHiC Environment Motivation• Concept • Analysis • SimANet CHiC - Chemnitz High Performance Computing Cluster - 512 AMD Opteron 64bit cores, 2.6GHz, each 4GB, fiber optics Infiniband network infrastructure - Scientific Linux 4.4, 2.6.18 Kernel Throughput measurements Pingpong measurements APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -11-

  12. Evaluation - SimANet Motivation• Concept • Analysis • SimANet • Simulation Platform for Ambient Networking • Java-based, complex simulation framework for heterogeneous MANETs / WSNs • Supports distributed computing environments •  Optimal testbed for Java-MPI parallelisation scenarios APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -12-

  13. Test Scenarios I Motivation• Concept • Analysis • SimANet • SimANet @ CHiC; up to 64 Nodes • Static node topology 100.000 nodes 5000 nodes APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -13-

  14. Test Scenarios II Motivation• Concept • Analysis • SimANet • SimANet @ CHiC; up to 64 Nodes • Dynamic node topology JMPI MPI2Java APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -14-

  15. Thank you … … for your attention. APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -15-

  16. Difference between MPI 1.0 & 2.0 Motivation• Concept • Analysis • SimANet MPI-1 - Point-to-Point operations - Collective operations - Synchronisation - Process groups … MPI-2 - Parallel I/O - Dynamic process management - Remote Memory Access (RMA) During the communication process, each command has to be called & executed in every communication node One communication node is able to initialise a communication process APCC 2010 Matthias Vodel, University of Chemnitz, Germany, vodel@cs.tu-chemnitz.de -16-

More Related