Date: Wed, 30 Aug 2000 16:57:17 +0200
From: Patrick Schaaf <bof@bof.de>
Without understanding the reason: all of the above is still valid,
but I found out that when I explicitly setsockopt(SO_SNDBUF) to 64k,
the writes become tame again.
David, is this intentional?
It means that this is how big your send buffers need to be to
optimally cushion your writer tasks from network and scheduling
latencies.
As discussed often on this list, calling setsockopt(SO_SNDBUF, N)
actually sets the socket buffer send buffer limit to 2*N because
of how we account the size of the skbuff structure _as well_ as
the data bytes against the socket.
But actually, in TCP nowadays the kernel should be determining this
value dynamically in response to network performance.
Alexey wrote all of the code which does this, so we'll have to wait
(about 1 week) for him to return and respond.
Meanwhile try tuning /proc/sys/net/ipv4/tcp_wmem to see if you can
obtain the same effect. TCP _DOES NOT_ use the
/proc/sys/net/core/*{r,w}mem* values at all anymore.
Later,
David S. Miller
davem@redhat.com
-
To unsubscribe from this list: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
This archive was generated by hypermail 2b29 : Thu Aug 31 2000 - 21:00:31 EST