This is the mail archive of the
systemtap@sourceware.org
mailing list for the systemtap project.
Re: static user probe performance
On 04/07/2010 07:43 AM, Frank Ch. Eigler wrote:
> (In this microbenchmark, it seems that the instrumentation is taking
> an impressive 98% of the cpu time, which ought to have triggered the
> overload protection. I assume there was an -DSTP_NO_OVERLOAD in
> effect.)
Well, the overload protection can only account for cpu time in our own
handler, and it's unfortunately blind to the overhead of the probe
mechanism itself. The int3 trap and all the uprobes code around it
(including the single-step) are outside the overload measurement.
Perhaps we should pad the measurements with approximate cycle counts of
a probe hit? Of course, people already complain when overloading kicks
in, and this would make it even more of a nanny...
Josh