Darshan 2.1.0 Release

This release primarily enhances portability and adds the option to use LD_PRELOAD for instrumentation rather than link time wrappers. This release does not add any new instrumentation or change the log file format.

Downloads Page

ChangeLog

  • additional environment variables to control log location, jobid and alignment parameters
  • additional configure tests to improve portability
  • bug fixes for darshan-parser –perf calculations
  • support for MPI1.x
  • support for OpenMPI
  • support for PGI and Intel compilers
  • new libdarshan.so dynamic library for use with LD_PRELOAD

Darshan 2.0.2 release

Changelog:

  • added a random identifier to job logs (to avoid collisions from multiple application instances within a single scheduler job)
  • improved installation and library path management for darshan-job-summary.pl
  • improved error handling in darshan-job-summary.pl
  • additional derived statistics categories for darshan-parser output:
    • ––all   : all sub-options are enabled
    • ––base  : darshan log field data [default]
    • ––file  : total file counts
    • ––perf  : derived perf data
    • ––total : aggregated darshan field data

Testpio Case Study #1

Last week I did some comparative runs of the “testpio” kernel to find out why pnetcdf I/O was slower than raw binary MPI-IO. In this scenario, 512 cores write a 51MB file ten times.

There were some minor differences:  binary (MPI-IO) uses a blockindexed type, while pnetcdf uses subarray.  Pnetcdf syncs the file a few more times – pnetcdf calls MPI_FILE_SYNC when exiting define mode, but I think we will change that soon.