2019.09.30 - ovis-hpc/ovis-wiki GitHub Wiki

• October 22-24, 2019
• Due to few presentation submissions we are looking at using 3rd day for one-on-one interactions with developers if desired
  • Assistance building, configuring, etc. on site resources
  • Troubleshooting
  • Discussions about v5 plans

• LDMS features:
  • Metric sets are now removed by ldms_set_delete
  • ldms_xprt_dir now conveys set meta-data information including size, and set_info information
  • libfabrics LDMS transport plugin
• LDMSD features:
  • ldmsd stream service:
    • A publish/subscribe service in ldmsd that allows external programs to send data (events) over an LDMS Transport to ldmsd plugins
    • Improvements to prdcr performance
  • ldms_ls provides summary set size information as an aid to ldmsd aggregator memory configuration
• New sampler plugins:
  • SPANK slurm_notifier: a Slurm SPANK plugin that uses ldmsd_stream to notify subscribers (plugins) of job events (e.g. start/stop).
    • Used by slurm_sampler, papi_sampler, and syspapi_sampler
  • slurm_sampler:
    • Multi-tenant capable slurm job information sampler
  • slurm_store:
    • SOS store plugin that converts multi-tenant job information into a form more suitable for analysis
  • PAPI Job Sampler (papi_sampler):
    • Collects hardware event counters per-process for all processes of a job
    • Receives configuration from a job's environment the slurm stream
  • PAPI System Sampler (syspapi_sampler):
    • Collects hardware event counters per-core, system wide
    • Uses libpfm for sampling and libpapi for event name to event-mask mapping
      • Allows consistent configuration to be used between syspapi and papi samplers.
    • Samples hardware performance counters on a per-core/uncore basis
  • papi_store:
    • SOS store plugin that converts PAPI job information into a form more suitable for analysis
  • IBM OCC sampler (ibm_occ)