On Mon, 2 Jul 2001, Brian Pane wrote:
> Here's a patch that disables the read before the select, but only for
> the HTTP client socket. Can somebody with a decent benchmark lab
> test it out to see if the throughput change is significant?
you can make some reasonable performance measurements with a pair of
machines and a crossover 100baseT segment. assuming you're able to max
out the 100baseT (shouldn't be hard) then you want to use "idle CPU" as
your metric. vmstat can help you there.
alternately you can use times() to report how much u+s time is consumed by
the processes...
also, copper gigabit cards are getting down to reasonable prices --
there's some $180 cards from 3com now. i think there are even cheaper
ones, maybe d-link? the 3com cheap ones are good hardware (with hw
checksum and scatter/gather), some of the others i doubt could ever run at
full gigabit speed, nowhere near enough buffering and no interrupt
coalescing.
btw, copper gigabit needs true 4-pair cat5 cables, and uses all 4 pairs in
a bidirectional manner... and you don't need to use a crossover at all if
you want to connect a pair of NICs back-to-back. it's really quite nice.
well except for the 15W heat-sinks on the big DSPs they need to do all the
signal processing :)
-dean
