Go to SCD News table of contents

Go to UCAR home pageGo to NCAR home pageGo to SCD home pageGo to SCD News home pageGo to SCD internal pagesGo to SCD News home pageGo to Features archiveGo to News archiveGo to Tips archiveGo to Updates archive SCD News > News item: May 29, 2001

Users benefit from blackforest "Mohonk" upgrade

Enhancements include GPFS compiling, LoadLeveler task geometry, addition of DPCL, and improvements to MPI, POE, and math libraries

blackforest

The IBM RS/6000, blackforest

 

by Richard Valent

On 18-19 May 2001, software engineers in SCD's Supercomputer Systems Group (SSG) upgraded the IBM SP blackforest to the "Mohonk" release as follows:

  • PSSP was upgraded from 3.1.1 to 3.2.
  • GPFS was upgraded from 1.2 to 1.3.
  • LoadLeveler was upgraded from 2.1 to 2.2.

The base AIX OS level remained the same. In addition to these changes, several patches were applied to blackforest's compilers and libraries. SSG was assisted in the upgrade by testing performed by software engineers in SCD's Technical Consulting Group.

The major advantages to users resulting from these upgrades are explained in detail in IBM document "PSSP 3.2: RS/6000 SP Software Enhancements." Here is a summary, with chapter pointers to IBM's documentation.

  • Ability to compile on GPFS (/ptmp) (chapter 4): You are no longer required to compile in a non-GPFS directory.

  • Ability to use LoadLeveler task geometry (chapter 5). LoadLeveler now allows you to group tasks of a parallel job to run together on the same node, possibly with a different number of tasks on different nodes.

  • Addition of DPCL (Dynamic Probe Class Library): A C++ based class library whose API enables a program to dynamically insert instrumentation code patches ("probes") into an executing program (chapter 6).

  • Enhancements to MPI (chapter 6):

    • MPI64. You can now compile MPI jobs with the -q64 option, thereby lifting previous stack memory size restrictions on the job.
    • Shared memory MPI. MPI programs may now benefit from using shared memory to send messages between tasks of a parallel job, running on the same physical node, without any modifications of original programs.
    • MPI-2 support. This adds new types of functionality to MPI according to the MPI 2.0 standard, including dynamic processes, one-sided communication, and parallel I/O.

  • Enhancements to POE (chapter 6):

    • Better and smaller (i.e. lightweight) corefiles
    • Improved parallel task identification

  • Routine and function enhancements to math libraries ESSL and PESSL (chapter 6):

    • ESSL now includes several new subroutines for solving dense symmetric indefinite systems, and a new linear least squares subroutine for general matrix factorization.
    • PESSL now includes several new subroutines for dense linear algebra and eigensystems analysis.

Please direct questions about this upgrade to the SCD Consulting Office (303-497-1278) or consult1@ucar.edu.

SCD News   ||  UCAR  ||  NCAR   ||   SCD   ||   Contact us   ||  Search
NCAR is managed by UCAR and sponsored by the National Science Foundation