Re: Performance analysis of Linux Kernel Markers 0.20 for 2.6.17

Hi Jose,

* Jose R. Santos ([email protected]) wrote:
> 
> The problem now is how do we define "high event rate".  This is 
> something that is highly dependent on the workload being run as well as 
> the system configuration for such workload.  There are a lot of places 
> in the kernel that can be turned into high event rates with with the 
> right workload even though the may not represent 99% of most user cases. 
> 
> I would guess that anything above 500 event/s per-CPU on several 
> realistic workloads is a good place to start.
> 
Yes, it seems like a good starting point. But besides the event rate, just
having the most widely used events marked in the code should also be the
target. The markup mechanism serves two purposes :
1 - identify important events in a way that follows code change.
2 - speed up instrumentation.

> 
> >On the macro-benchmark side, no significant difference in performance has 
> >been
> >found between the vanilla kernel and a kernel "marked" with the standard 
> >LTTng
> >instrumentation.
> >  
> 
> Out of curiosity,  how many cycles does it take to process a complete 
> LTTng event up until the point were it has been completely stored into 
> the trace buffer.  Since this should take a lot more than 55.74 cycles, 
> it would be interesting to know at what event rate would a static marker 
> stop showing as big of a performance advantage compared to dynamic probing.
> 

In my OLS paper, I pointed out that, in its current state, LTTng took about 278
cycles on the same Pentium 4. I think I could lower that by implementing per-cpu
atomic operations (removing the LOCK prefix, as the data is not shared between
the CPUs; the atomic operations are only useful to protect from higher priority
execution contexts on the same CPU).

Regards,

Mathieu

OpenPGP public key:              http://krystal.dyndns.org:8080/key/compudj.gpg
Key fingerprint:     8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68