|
Hi,
I want to measure the overhead of LVS, e.g. the additional latency imposed
by LVS instead of a traditional router. I looked up several documents
and found Zhang's paper mentioned 60us for rewriting a 536 bytes packet.
I reckon this was with NMI disabled and no tasklets. Actually how can
one explain 60us? This is rather nothing.
All measurements in other documents focuses on throughput instead of
latency overhead.
Well, I think this is because the network is already fast enought to
process a big enough number of packets for most people so the loss in
packets because of the overhead is marginal.
I believe it's not very hard to measure the latency as long as we have a
fine grain clock. Anyone knows how to get fine grain time?
You might wanna check out the following sites as a starter:
http://oss.sgi.com/projects/kernprof/
http://oprofile.sourceforge.net/about.php3
Make sure you enable profile support when you boot the linux kernel. If
you're not lucky with this approach, ping the IBM guys doing the linux
scalability effort at http://lse.sourceforge.net/. I'm sure they can
help you better on that issue. Just make sure you come back to this list
with your results ;).
Best regards,
Roberto Nibali, ratz
--
echo '[q]sa[ln0=aln256%Pln256/snlbx]sb3135071790101768542287578439snlbxq'|dc
| 