On Jan 30, 2009, at 6:29 PM, jared r r spiegel wrote:
On Fri, Jan 30, 2009 at 05:55:48PM -0800, Philip Guenther wrote:
It seems the
only significantly value for net.inet.tcp.rfc1323 is 0 (disabled)
vs. non-0
(ws=0). Am I missing something?
You'll never see a scale size larger than zero unless the involved
program sets a socket receive buffer size larger than 64KB before
calling listen() or connect(), that being the value from which the
receive window size is derived.
so in other words, if you want wscale to be able to go to 1 but for
things
who don't support wscale you want to retain the same current
functionality,
add 65536 to the current value of whatever (send|recv)space you're
talking
about.
wscale of 2? add 131072 from the baseline, etc.
go too high and stuff won't work at all
--
jared
Great, thanks for the pointers! I'm trying to fiddle with iperf
performance testing going to a Linux box. tcpbench works great on
OpenBSD, but it seems iperf is the only thing readily available for
Linux that is also on OpenBSD. I'm just trying to figure out how each
variable influences the throughput.
--
bk
