Date: Fri, 24 Jan 1997 13:53:44 +0900 From: Kenjiro Cho <kjc@csl.sony.co.jp> To: hackers@FreeBSD.ORG Subject: tcp window size problem, again Message-ID: <199701240453.NAA12432@hotaka.csl.sony.co.jp>
next in thread | raw e-mail | index | archive | help
Recently, I did some network performance benchmarking on high speed
network (ATM and 100baseT), and found that the FreeBSD's score is far
behind other systems.
It turns out that the TCP window size gets reset to the default value
at a connection establishment on FreeBSD. After digging into the
hackers ML archive, I found this problem was already discussed back in
September. (see Garrett's comment attached below.)
I don't have a good solution to resolve the conflict between a value by
SO_ SNDBUF/SO_RCVBUF and a value by the route. But at least we should
leave out setting stupid rmx_snedpipe/rmx_recvpipe values.
The current in_addroute() in in_rmx.c sets tcp_sendspace/tcp_recvspace
to rmx_sendpipe/rmx_recvpipe. This has no effect when the socket
buffer size is not changed, but it does override the socket buffer
size even when the socket buffer size has been changed.
There is no positive effect to set these default values.
By leaving out the default values for rmx_sendpipe/rmx_recvpipe,
the behavior of TCP falls back to the original BSD implementation.
- SO_SNDBUF/SO_RCVBUF work as expected.
- administrators can override the socket buffer size by explicitly
setting sendpipe/recvpipe to a route.
We can leave the problem of the conflict until someone comes up with a
better handling of rmx_sendpipe/rmx_recvpipe.
Anyway, with a larger socket buffer size, the TCP score jumps up from
70Mbps to 120Mbps! It's pitty if FreeBSD gets a reputation of bad
netwrok performance by benchmark tests.
--Kenjiro
----------------------------------------------------------------------------
Date: Thu, 12 Sep 1996 15:05:04 -0400
From: Garrett Wollman <wollman@lcs.mit.edu>
To: Luigi Rizzo <luigi@labinfo.iet.unipi.it>
Cc: hackers@freebsd.org, wollman@lcs.mit.edu, phk@freebsd.org
Sender: owner-hackers@freebsd.org
Subject: Changing default TCP window size ?
<<On Wed, 11 Sep 1996 23:17:07 +0200 (MET DST), Luigi Rizzo <luigi@labinfo.iet.unipi.it> said:
> How do I force TCP (or the OS) to use a window size different from the
> default 16K ?
> The following code
> mywin=DESIRED_WINDOW_SIZE;
> setsockopt(f, SOL_SOCKET, SO_SNDBUF, (char *)&mywin, sizeof mywin)
> appears to set the value correctly (it can be read back with
> getsockopt) but then it appears not to have any effect on TCP
> (monitoring the connection with netstat -na always shows a Send-Q
> around 16KB, and the throughput in window-limited configurations
> remains unchanged).
This is a conflict that needs to be resolved between the
option-processing code and the code that attempts to set the socket
buffer size based on what's in the route. I can't give you a
fully-fleshed-out solution, but essentially the code needs to check if
there is a buffer size specified via setsockopt, and if there is, then
it should ignore the values in the route (or perhaps it should use the
lock bits in the route to indicate some sort of clamping). I have a
long-standing bug report form Matthew Dillon about this, but it has
not had a sufficiently high priority to get done. Come up with a
good fix and I'll take it.
-GAWollman
--
Garrett A. Wollman | O Siem / We are all family / O Siem / We're all the same
wollman@lcs.mit.edu | O Siem / The fires of freedom
Opinions not those of| Dance in the burning flame
MIT, LCS, ANA, or NSA| - Susan Aglukark and Chad Irschick
----------------------------------------------------------------------------
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199701240453.NAA12432>