Hi All,
As you know, during the Bake-off planning meeting, we decided to
try to have low-level, per-packet delays on the polygraph clients
and servers. We proposed to use the "DUMMYNET" feature of FreeBSD
to simulate such delays.
We have been working hard the past couple of weeks with the dummynet
configuration. Our goal is to find out what request rate we can
sustain with Polygraph and various dummynet configurations. After
all these tests, we are forced to conclude that the dummynet code
in FreeBSD 3.1 is not suitable for simulating per-packet delays
with the PolyMix-2 workload. We observed a high amount of variation
in actual packet delays. This, of course, leads to an undesirable
and unreproducable workload.
We are now faced with choosing from one of the following three
options. We need your input on this matter!
1) Drop dummynet (and per-packet delays) entirely. We would have
the same server-side "think-time" delays that were used for
PolyMix-1 and Datacomm-1 workloads. Taking this option, which
eliminates client-side delays, means there is no difference between
our proposed "ISP" and "Corporate" configurations (ref "Overview"
in http://bakeoff.ircache.net/N02/tests.html).
2) Dummynet *may* be able to support a lower request rate per
client/server pair. That is, instead of 400 req/sec per
pair, we may be able to have 200 req/sec. This would double
the number of PC's needed during the bakeoff, and increase
your participation costs.
3) FreeBSD version 3.3 may have improved dummynet code that
works well with Polygraph at 400 req/sec. We have already
committed to using FreeBSD 3.1 for the tests, however. We have
been in contact with the dummynet author, and he believes the
3.3 code is good enough.
At the moment, our default option is #1. We are skeptical that
Dummynet is currently stable enough to be used in a benchmarking
environment. However, we are also investigating how well it
works in the FreeBSD 3.3 kernel. Options 2 and 3 may also
work okay. In any case, we need your input! Please let us know
what you think.
Duane W.
This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:09 MDT