[Click] NAT Latency Measurement
Latency Buster
latencybuster at gmail.com
Fri Dec 18 12:41:24 EST 2009
> Do you really care about latency? Or just overall performance (which is
> limited by the sum of the per-element latency)?
In the end I care about port to port latency which is the sum across
all the elements. However, I would like to get a feel of the most time
consuming element in the path so that I can tweak it for the last
drop.
> Are these packets unique already? or is expensive_uniquify being called?
No expensive uniquify calls. Checksum check is false too.
I have taken the measurements when the mappings have been installed.
i.e - at least one packet of the flow was seen. I have used user level
click with file being read using FromDump() and did a performance
measurement - that figure comes to less than 20us. The kernel level
version of click with live traffic should at least beat that figure.
I think I will do a oprofile measurement and see how the cycles are
being consumed. In the meantime, if you guys think that there is a
better/different approach to break down the latency that woule be very
much appreciated.
More information about the click
mailing list