Re: Zipf vs. Uniform

From: Alex Rousskov (rousskov@ircache.net)
Date: Mon Dec 06 1999 - 17:19:44 MST


Greetings,

        We have done more experiments with Zipf and Uniform. I am almost
convinced now that Zipf(0.4) and Zipf(0.6) will produce unrealistic skew
in the access patterns if used with Polygraph. For some of the results,
please see the table and graph at

    http://polygraph.ircache.net/Workloads/PolyMix-2/#Sec:WorkSet

but read all the comments under the table before making your
conclusions.

We could still go with Zipf(0.1) or Zipf(0.2) because these
distributions are probably close to Uniform. Again, this approach sounds
like a placebo to me. Besides, we need to leave you something to
complain about.

As you know, there was virtually no public _data_ posted to our lists
that would help us to make the choice. We got one private message saying
that Polygraph memory hit ratios (with Uniform and PolyMix-2) are lower
than the real memory hit ratios, but due to other specifics of the
workload, going with Zipf would make the load "too easy" on a cache. The
vendor claimed that the workload is already ~50% easier than real
traffic. A few other messages voted for Uniform because (in authors
opinion) it is better to have a tougher workload than an "easier" one.
Clearly, all these arguments cannot be used directly to decide on Zipf
value.

If we get no more information from you, we intend to preserve Uniform
popularity model for PolyMix-2. Again, no changes to worry about.

Alex.

P.S. BTW, has anybody tried to calculate Zipf's alpha parameter for
     a log produced with Uniform Polygraph workload? (would be a good
     illustration for the "everybody must use Zipf" argument :) )



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