[ltt-dev] performance results I measured with the latest lttng patches

[Mathieu Desnoyers]

* Michael Rubin (mrubin at google.com) wrote:
> On Wed, Sep 24, 2008 at 12:25 AM, Jan Blunck <jblunck at suse.de> wrote:
> > Isn't lttng-compile-out missing here? Or did you ran the same base kernel?
> 

Hi Michael,

> I believe it's the same base kernel. Which means we have about a 2.3%
> performance penalty for just compiling the LTTng code in with no
> markers enabled. We love the functionality of LTTng, but that is a
> non-starter for us.  We are motivated to help make this better.
> 

Hrm, this result surprises me.

> I am curious what the expectations and goals the lttng efforts have
> for performance in these scenarios. Do you guys have a goal for perf
> impact in the situations of not configured (or not compiled) vs
> configured (compiled in) but no trace points enabled vs tracepoints
> enabled?
> 

yeah, about, say.. unnoticeable ? :)

> What sort of benchmarks are you using to test performance? That way we
> can use the same ones to better compare notes. Do you have access to
> HW with lots of cores?
> 

A while ago I used a mix of lmbench and my own microbenchmarks (running
markers and tracing in tight loop in a kernel module to have an idea of
the cache-hot behavior). I've recently seen from Jiaying posts that
tbench does a very good job at stressing LTTng, given it uses system
calls extensively.

I have a dual-quad cores at my lab, I'll run some tbench tests on this
setup to try to figure out what went wrong. I guess there might be some
option enabled when it shouldn't, or maybe I made some sort of stupid
error.... :) Anyway, that should not be too hard to find, hopefully.

Mathieu

> Thanks in advance,
> 
> Michael Rubin
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68