clustering

John Snell (geigudr@cs.washington.edu)
Fri, 29 May 1998 15:11:38 -0700 (PDT)

Some of those who got to the meeting early heard a part of the problem
encountered in the measurements: DNS/bind is a serial application, and
URL's are typically DNS names. The solution of using the IP address, for
some reason, doesn't always work -- there are at least two examples of
URL's where the IP, although reported as being the address of the machine
referenced by the WWW prefix, just doesn't work.

So, back to DNS resolutions. Bind is serial. Which means that there's a
bottleneck in the measurements -- everyone queues up behind the BFE server
that takes 15 minutes to resolve, when multiple traces are to be
performed.

Originally, this meant that they all timed out, and the results were not
kept -- BFE's revenge.

Currently, I've got it set up to allow DNS resolutions to timeout at the
Java layer (there's some hidden internal timer), and then our timer after
that. Which means that all the requests so queued become synchronized to
the resolution of BFE. Of course, their new wait times are still
according to the Uniform distributions, once this trace is complete.

SO. Is this synchronization a problem? Should we drop the servers that,
for some odd reason, don't resolve properly when using IP's, or should we
allow that synchronization to occur?

Although I don't like the synchronization, I similarly don't like the idea
of removing more datapoints.

And one last question. Was the twonk who originally wrote bind working in
MS-DOS, or some similarly-valued single user system? This lack of
parallelism is annoying.

Thanks,
John

_____________________________________________________________________________
"The human mind is a 400,000-year-old legacy application...and you expected
to find structured programming?" -- Randall Davis, 1996 AAAI Pres. Address