Subject: A bug in the Congestion Window Monitoring code?
To: None <email@example.com>
From: Tom Ivar Helbekkmo <tih@kpnQwest.no>
Date: 09/26/2001 13:30:27
On the 10th of September, congestion window monitoring was enabled by
default in -current (net.inet.tcp.cwm=1). I've just realized that a
problem I've observed lately, and finally found time to look at today,
was caused by this.
The file systems on my workstation (running NetBSD/i386-current) are
backed up to a FreeBSD 4.2 system using Amanda. Since I upgraded to a
version of -current that has CWM turned on, performance has been
abysmal; the two machines connect to the same 100Mbps network, and the
dumps were transmitted at a steady rate of slightly more than 60Kbps.
Using tcpdump to observe the data stream, it was very regular: a burst
of four 1460 byte packets of data from my machine to Amanda, then a
100ms pause, then an ack from Amanda to me, immediately followed by a
new four packet burst, 100ms pause, ack, burst, pause, ack, burst...
After observing this, a colleague and I took a quick peek at sysctl
variable values in the net.inet.tcp section -- and observed that
net.inet.tcp.cwm_burstsize was set to 4. This looked so likely, that
we immediately tried changing net.inet.tcp.cwm from 1 to 0. The xterm
that I was running tcpdump in promptly turned into a gray blur. :-)
I'm leaving it off for now, but if there's anything I can do to help
debug the situation, I'd be happy to. It's perfectly reproducable, on
a daily basis...
Puritanism -- the haunting fear that someone, somewhere, may be happy.