persistent tcp connection?

Fran Lawas-Grodek Fran.Lawas-Grodek at
Wed Mar 26 18:48:27 PST 2003


Hopefully someone has some advice for us.

We are planning to test a modified version of the tcp stack
under a high delay link with a specified bottleneck of 5Mbps,
but first, we are trying to see how the normal tcp stack performs
under the same conditions.

When we send ttcp transfer tests back-to-back with a 5 minute
wait in between, the first transfer gives us the expected
throughput with the expected retransmissions.  However, the
subsequent transfers do not show any retransmissions and the
throughput is much slower, as if something is compensating for
the previous retransmissions.  We would like to see a performance
in the following transfers similar to the first transfer.  If
we wait 2.5 hours between transfers or even if we reboot the 
ttcp systems between each transfer, then we would see a similar
throughput complete with the expected retransmissions,
but we'd prefer to do neither for in favor of a more timely

Does anyone know if there a sysctl parameter or kernel option that 
would "clear out" any memory of a previous tcp connection?  We've
played with the following sysctl parameters, and these are what
they are currently set --

net.inet.tcp.keepidle: 14400
net.inet.tcp.keepintvl: 150
net.inet.tcp.always_keepalive: 1

Any other thoughts?

This is on a pc running FreeBSD 4.1

Your advice would be greatly appreciated.

Fran Lawas-Grodek

Frances J. Lawas-Grodek       |     
NASA Glenn Research Center    |     Fran.Lawas-Grodek at
21000 Brookpark Rd, MS 142-4  |     phone: (216) 433-5052
Cleveland, Ohio  44135        |     fax  : (216) 433-8000

More information about the freebsd-net mailing list