Re: Weird tcp performance differences with 2.0 and 2.2 kernels

Harvey J. Stein (hjstein@bfr.co.il)
10 Feb 1999 21:19:07 +0200


Andi Kleen <ak@muc.de> writes:

> In article <m2679aio4w.fsf@blinky.bfr.co.il>,
> hjstein@bfr.co.il (Harvey J. Stein) writes:
> > "Mark R. Boyns" <boyns@contigo.com> writes:
> >> We're trying to figure out some weird performance differences between
> >> 2.0.x / 2.2.x kernels and the interface being used.
>
> > I noticed some weird behavior last night. I was dialed into the
> > office from home. I was uploading some files via ftp. The modem send
> > light started off blinking rapidly (basically flickering) indicating
> > high traffic. Then it suddenly backed off, maybe sending about 1
> > packet/second. It stayed like this for maybe 30-60 seconds & then
> > sped up again, but not to as fast as it was initially. Final
> > throughput was a little on the low side.
>
> Congratulations. You just discovered the slow start TCP algorithm. Please
> read RFC2001 for more information.

It's not explained by slow start. Ping times are <200ms, so even with
the congestion window reduced to size 1 my machine still should have
only been waiting about 100ms for an ACK before transmitting the next
packet.

It could be congestion avoidance, but I fail to see why this should be
triggered when I'm uploading over a 33.6kbps ppp link to a machine on
an ethernet. It's transfers in the other direction that should
trigger congestion avoidance.

In fact, what seems to be happening is that eventually it's settling
on a congestion window size of 1, waiting for each ack before sending
another packet. This basically agrees with what I observed after it
sped up again, but still doesn't explain the period of time where it
only was sending about 1 packet per second. It also doesn't explain
why it didn't increase the congestion window during this.

BTW, I only brought it up because I haven't observed this behavior
with 2.0.x.

-- 
Harvey J. Stein
BFM Financial Research
hjstein@bfr.co.il

- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/