On Tue, 30 Nov 1999, Pei Cao wrote:
> I am going to shamelessly quote my own paper here :-) Our INFOCOM paper
> (http://www.cs.wisc.edu/~cao/papers/zipf-implications.html)
> calculated the alpha value for UC Berkeley traces, DEC Corporate traces,
> as well as a few others. The values range from 0.64 to 0.83;
...
> I think 0.6 is a good choice. If one is really concerned about high memory
> hit ratio, I would say go with 0.4, which are below 90% of the data points
> on the curve at http://www.ircache.net/Cache/Statistics/Popularity-Index/.
OK, we are finally moving somewhere. So the "real" values for alpha are
probably somewhere in the 0.6 - 0.9 range and may be decreasing. Polygraph
workload is not "real" so some adjustments may be necessary, especially
keeping in mind the fact that none of the studied traces were close to the
request rates some vendors are pushing through their boxes now.
I suggest that the bake-off vendors and anybody else who is interested try
two settings for the popularity model: Zipf(0.4) and Zipf(0.6), compare
them to the current uniform model, and let us know which setting produces
more realistic workload (in terms of memory hit ratio or any other metric
that you prefer).
To try these, all you need to do is to edit polymix-2.pg and change
pop_model = pmUnif();
to
pop_model = pmZipf(0.4);
or
pop_model = pmZipf(0.6);
You can try other values as well, of course. We can probably decide on the
final value for the parameter as soon as we get five or so responses.
Thanks,
Alex.
This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:10 MDT