Friday, September 19, 2008

Timed measurements with pmcstat(8)

Every once in a while I get email asking for an option to pmcstat(8) that would allow hardware events to be measured for a specified time interval.

The good news is: pmcstat(8) already supports timed measurements — using /bin/sleep.

Here is how you would do it:

  • Timed system sampling would be done in the following manner:

    % pmcstat -S instructions -O logfile /bin/sleep 42

    This invocation allocates a system-scope sampling PMC (-S) and profiles the whole system while /bin/sleep executes, i.e., for 42 seconds.

  • Timed measurements on groups of processes are performed in a similar fashion:

    % pmcstat -d -p dc-misses -t '1234' /bin/sleep 24

    This invocation would allocate a process-scope counting PMC (-p) that counts data cache misses, attach it (-t) to the process with pid 1234 and its descendants (-d), and count for 24 seconds.

    Note that the '-t' option also takes regular expressions, so you don't need to know process ids beforehand. To count instructions executed by processes named 'httpd' for an hour you would use:

    % pmcstat -p instructions -t 'httpd' /bin/sleep 3600

The "Unix way" uses small tools, each of which does a defined task reasonably well and which are combined to perform more complex tasks. In the examples above, /bin/sleep manages time intervals and pmcstat(8) manages PMC based measurements.

When you combine them, voilà, you get timed PMC based measurements.

Monday, September 15, 2008

PMCTools meets KCachegrind

Fabien Thomas sent in a screenshot showing the result of his experiments at integrating KCachegrind and PmcTools.

Thank you Fabien, for sharing!