On Wed, Mar 08, 2006 at 06:24:02PM +0100, Marcel Prisi wrote:
>
> I read some old threads about too small tcp.sendspace / tcp.recvspace in
> 3.4 time that used to hit performance so I thought it would be useful.
of all the times i dicked with those, my results have been that
any performance gain (*if* any) i get in throughput for long-lived connections
which ride on an essentially trouble-free link between me and the
other host i'm testing on are offset by an increased latency i
see between me and my default gateway.
shorter connections, i've found, tend to appreciate not increasing those
values.
eg, i start a ping to default my gateway, observe the RTT; observe
my tcp sysctl crappers, start a large tcp xfer to someone else in town
who is at most one router (and a few ATM hops) away; observe
the throughput of the xfer and observe the current RTT to the gateway
while that is happening; dick with sysctl {send,recv}space; repeat.
don't have hard records of the results, but iirc, where normally my
RTT is on the order of 12ms, with default tcp sizes, under full
congestion (tested up and down streams individually) i was seeing
something like a jump to 600ms while downstream saturation and
2200ms while full upstream saturation; jacking the tcp to 64k like
everyone uses increased the throughput of the transfer by some 8-15%,
but sent the RTT to something like 1500ms/8000ms respectively.
ALTQ was not in use.
one may say 'tcp window size has nothing to do with ping times',
but in my non-expert experience it certainly does when the pings and
the tcp are sharing the same congested pipe; or at least when an
encapsulation from IP to ATM (dsl) is involved.
there may be other factors too, and i am actually at a loss to explain
it beyond the conclusions that one can draw based on watching the results,
but the delightful thing is it doesn't particularly matter why to me
since the default settings have proven themselves (to me) to be just
fine in all cases (of my connection), and thus far i've found that
changing them usually degrades one trait of my link more than it improves
another one that i value less.
profile at the dslam is something similar to 3Mb down / 768Kb up, interleaved.
--
jared
[ openbsd 3.9-beta GENERIC ( jan 30 ) // i386 ]
