Correlating timestamps in user and kernel space

This proposal has been accepted as a session.

*
Accepted Session
Microconference Session

One Line Summary

We need a way of timestamping perf(ormance) data generated in userspace.

Abstract

There are many great sources of performance data in the kernel, like perf and ftrace. Unfortunately anyone trying to correlate data coming from them with a userspace event, for example JIT compiler generated debug symbols, is up for a disappointment. The main is lack of common “time zone” between kernel usually using “cheap” time sources like sched clock, and userspace-available time of day or POSIX clocks.

There has been numerous approaches proposed on kernel mailing lists over the last year or two, from extra trace events, through custom ioctls (http://article.gmane.org/gmane.linux.kernel/1433933) which was quietly ignored, to specialised POSIX clocks (http://thread.gmane.org/gmane.linux.kernel/1376353/focus=1524878) which was rejected. Ingo suggested synchronising sched clock in sync with the raw monotonic one, so I came up with a simple proof of concept (http://article.gmane.org/gmane.linux.kernel/1751163). Some liked the basic idea, Richard polietly suggested getting back to the trace events… History repeats itself ;-)

The subject comes and goes for the last two years (at about a half-yearly intervals), so it would be great to be able to discuss the thing face to face and (maybe?) come up with some decisions…

Tags

perf, ftrace, time

Speaker