UPC at LBNL/U.C. Berkeley Overview - PowerPoint PPT Presentation

1 / 3
About This Presentation
Title:

UPC at LBNL/U.C. Berkeley Overview

Description:

Enable wide-spread use of UPC by providing a portable open source compiler for UPC ('gcc-like' ... Based on Open64 open source C and Fortran compiler ... – PowerPoint PPT presentation

Number of Views:20
Avg rating:3.0/5.0
Slides: 4
Provided by: gab143
Category:
Tags: lbnl | upc | berkeley | open | overview | spread | wide

less

Transcript and Presenter's Notes

Title: UPC at LBNL/U.C. Berkeley Overview


1
UPC at LBNL/U.C. BerkeleyOverview
  • Kathy Yelick
  • U.C. Berkeley, EECS
  • LBNL, Future Technologies Group

2
Unified Parallel C (UPC) at LBNL/UCB
  • Unified Parallel C is a parallel language
    designed to
  • Make parallel machines more accessible to users
  • Enable better efficiency and scalability than
    other models
  • Based on research languages from Berkeley, DoD,
    and DOE labs
  • UPC has broad support
  • UPC compilers available on Cray, SGI, and HP
    machines
  • Over 100 attendees at the last UPC meeting
  • The Berkeley UPC Compiler is a joint LBNL/UCB
    effort
  • Lead by Kathy Yelick, with 4 LBNL staff, 2 UCB
    students, and 2 both
  • About 1M in annual funding (through the DOE
    Programming Models Center, DOE base program
    funding, and DoD funding)
  • Goals
  • Enable wide-spread use of UPC by providing a
    portable open source compiler for UPC
    (gcc-like)
  • Identify and demonstrate optimizations for this
    class of languages, which are relatively
    unstudied in the compiler community

3
Berkeley UPC Compiler Results
  • First release scheduled for this month, March 03
  • Based on Open64 open source C and Fortran
    compiler
  • Generates C as portable assembly language and
    one-sided communication a layer called GASNet,
    developed by the group
  • Tuned networking layers for Myrinet, IBM-SP, and
    Quadrics, with reference implementation for MPI
  • Initial benchmarks show
  • In spite of portability constraints, such as
    generating C, the Berkeley compiler often beats
    the HP compiler on HP hardware!
  • Performance is comparable to MPI on the NAS
    Parallel Benchmarks (within a few percent)
  • Several optimization opportunities exist
Write a Comment
User Comments (0)
About PowerShow.com