Metrics Required for the performances tests

Metrics from readout performance test template

Copy of version: https://docs.google.com/document/d/1LoKOntjo5ESYEkcmfScwtCp7qwCInjv2G0lgnaopAXI/edit

Failure criteria and potential metrics to identify them:

• Data reception failures
  • missed and dropped packets
• Trigger primitive generation failure
  • post processing error messages
  • TP rate per channel
• SNB recording failure
  • error messages
  • data written for one APA must be ~876.25 GB
  • data written for one CRP must be ~1024 GB
• Excessive system utilisation:
  • CPU utilisation for all cores must be less than 80% (only average available)
  • memory bandwidth across each NUMA nodes must be less than 80% (unclear which numa node is presented)

Technical specifications of the machine tested must be logged in each report:

• Motherboard and chipset
• CPU details (incl. caches)
• DRAM (incl. speed, capacity, and available channels)
• NICs (incl. for readout and backend with available ports, bandwidth and features.)
• SNB store (incl. drive specs. and their configuration -RAID-)
• OS and software (OS, kernel version, external dependencies used)

Concurrancy:

• Active user sessions on the RU during the load test
• Active processes and services running on the RU during the load test (top ten dominant processes, usage of each daq application on the RU)

Post test Analysis:

• Overall RU utilization in terms of:

  • CPU instructions per second and percentile per core
  • Memory BW utilization per channel (what are channels in this context?)
  • CPU cache utilization in terms of misses and hits

• Every application and their threads’ utilization in terms of:

  • Peak and average CPU percentile
  • Peak and average memory BW utilization
  • Peak and average CPU cache utilization in terms of misses and hits

Additional information in reports:

• Kernel’s and other services’ resource utilization
• Available headroom on the RU, in terms of:
• Unused cores
• Unused memory bandwidth
• Downstream network bandwidth available
• Configuration information for readout
• Configuration information for trigger primitive generation
• CPU pinning information

What do instructions per cycle and any retired tell us? Why is metrics split in Socket number rather than numa node?