Experiences with kprobes

[Baruch Even]

Frank Ch. Eigler wrote:
> Hi -
> 
> On Tue, Mar 22, 2005 at 03:13:17PM +0000, Baruch Even wrote:
> 
>>[...]  In the test with the probes I only get the cwnd up to about
>>3000, with a round trip of 300ms and delayed acking, it means
>>3000*1000/300/2 = 5000 packets per second. [...]  we get double the
>>rate of ACKs which sends us to 10000 pps, at which stage the probes
>>will probably affect behaviour significantly enough. [...]
> 
> But 10000 packets per second at 10000 clocks per packet (from your
> previous message) represents only 1/30th of the raw capacity of a 
> 3 GHz CPU.  Before instrumentation, it would have been 1/60th or so.
> It's strange that this 1/60th difference is enough to bog your 
> test down.

In my updated testing of the fully optimized code of 2.6.6 it still gets 
the work done. On the partial optimization done for 2.6.11 it performs 
worse than the same optimizations for 2.6.6 with the old instrumentation.

So at the end of my porting work to 2.6.11 everything will work, but I 
am hoping to show the step-by-step improvements and not just the jump 
from start to finish. I also need to justify each patch to the LKML 
crowd, and to myself.

And the target goal is to be able to work at least for 1Gbps rate and 
I'm only at a third of that...

So yes, I can use Kprobes for now but I'm looking for something with 
less overhead if I can find it. I do appreciate the work you are doing 
with kprobes, it looks (and is!) amazing, but I'm not sure this is the 
best tool for my task at this time.

Baruch