Systemtap effect on performance

Question

I'm looking for some real information about the effect of systemtap on performance. I know that static kernel probes are documented somewhere on their website (with stats), but what about userspace probes?

I saw this post on their mailing list (http://sourceware.org/ml/systemtap/2009-q1/msg00518.html) which points at a possibility of single-steping in some cases. Can anyone explain / point at some resources that list which techniques are used for which probes?

What overhead will process().function() introduce? What about static marks?

Answer 1

Here is some benchmarks of the overhead on a ppc64 machine. This is all I could find. It appears that SystemTap was designed to be a very low overhead tracer according to this article.

It would seem that the utrace and uprobe user-space side of SystemTap is not where most of the work is done. So it shouldn't be too much different from the kernel trace benchmarks given. Here are some utrace performance tests.

Since the current version is only 0.8 I guess there has not been much work on quantifying the performance and overhead of SystemTap.