Re: input needed on packet delay issues

From: Duane Wessels (wessels@ircache.net)
Date: Wed Dec 01 1999 - 17:40:00 MST


Hello again...

I'll bring you up to date on the packet delay issues....

To summarize, I said that Dummynet was not working well with PolyMix-2
and our target request rate. It is unstable and introduces too much
uncertainty into a test. I mentioned three options:

        1) give up on dummynet and low-level packet delays
        2) try dummynet at lower request rates
        3) try dummynet on FreeBSD-3.3

Most people preferred #3, then #2, then #1.

We have been running our tests here with FreeBSD-3.3. We ran a 16-hour
long no-proxy test between one client PC and one server PC test at 700
req/sec. These results look much better, but there is still some
instability. Instead of getting a relatively constant response time,
it varies betweeen 3850 and 4075 msec. So there are two issues: (a)
the "absolute value" is higher than what we predict, and (b) there
is too much variation. You can see these results at
http://www.ircache.net/~wessels/Junk/clt.700.dn.png

We are now running additional tests with lower request rates (400/sec),
and some without dummynet to determine its effect. I'm optimistic
that 400/sec will give us good results.

Also note that, on the advice of the dummynet author, we changed
FreeBSD's clock resoltion to 1msec. The default was 10msec.

FYI, here are the commands that we use to configure dummynet
on FreeBSD:

        ipfw -f flush
        ipfw pipe 1 config delay 75ms plr 0.001
        ipfw pipe 2 config delay 75ms plr 0.001
        ipfw add pipe 1 ip from any to 10.8.0.0/16 in
        ipfw add pipe 2 ip from 10.8.0.0/16 to any out

On the clients we're using 75msec delays and 0.1% packet loss.
On the servers we use 55msec delays and 0.05% packet loss.

Duane W.



This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:10 MDT