Perf tool data with perf-archive captured can be HUGE

[portante]

The perf tool data collects a perf-archive as part of the tool's stop operation. This archive collection can result in a HUGE amount of repetitive data being collected.

It was observed in one run that a single perf.data.archive.bz2 file was 147 MB in size. With two nodes collecting perf tool data, across 20 iterations, with 3 samples in each iteration, the combined size was a whopping 17 GB (147 * 2 * 3 * 20 = 17,640 MB).

There are a few options we might want to consider:

• Adding a perf tool option to not capture perf-archive output
  • and turning off the capturing of the perf-archive output by default
• Adding a way to allow the user to remove the perf.data.archive.bz2 files after a pbench benchmark finishes, but before the user executes pbench-move-results
  • This would allow a savvy user a chance to pick which iteration(s)/sample(s) they would want to keep, and for which nodes
• Another option would be to use multiple tool groups, collecting samples for each
  • Perhaps pbench-uperf --tool-groups=light,heavy --samples=4,1 would run 4 samples using the light tool group, and one sample using the heavy tool group, where the perf tool might be in the heavy but absent from the light
    • In the example above, if each iteration had only 1 sample collecting using the perf tool, then you'd have 5 GB (147 * 2 * 20 = 5,880 MB) of data collected