Re: Repeated simulations

From: Alex Rousskov <rousskov_at_measurement-factory.com>
Date: Thu, 25 Oct 2001 10:07:54 -0600 (MDT)

On Thu, 25 Oct 2001, Svensson, Andreas wrote:

> If I split the polymix-3 simulation into two parts (fill and
> test), I don't get the same results as with regular simulations
> when I have heavy traffic. E.g. for 400rps:
>
> My test after fill: Response time: 1801.28 msec - misses: 3211.60
> msec - hits: 241.00 msec Hit Ratio: 50.52 %
>
> Regular Polymix-3: Response time: 1858.70 msec - misses: 3214.92
> msec - hits: 152.50 msec Hit Ratio: 46.93 %
>
> For low traffic, e.g. 150rps, I don't have this problem. Why do I
> get different results when I split the simulation?

Unless specifically configured to do so[1], Polygraph does not use
URLs from one test in another test. Thus, when you split fill and
measurement phases into two tests, your measurement phases have to
build working set from scratch.

Moreover, you need to make sure that when you split fill and
measurement phases, your working set size does not grow indefinitely
during the measurement phases. The latter may lead to worse hit
ratios, especially for higher request rates when the cache is less
likely to be able to keep the entire faster-growing working set.

> Is there any way I can fix that?

If the split is done correctly (which means you would still need to
build and freeze your working set before measurement phases), the
performance difference should be marginal, but you will be losing time
unless you run one fill test and many measurement tests. That is, the
total (fill test + measurement test) duration will exceed the ((fill +
measurement) test) duration because of that extra time to rebuild the
working set.

It would be great if one could seamlessly reuse the working set from
previous tests. The "persistent working set" feature is on our wish
list[2].

Alex.

[1] http://www.web-polygraph.org/docs/userman/repeat.html
[2] http://www.web-polygraph.org/support/wishes.html
Received on Thu Oct 25 2001 - 10:11:07 MDT

This archive was generated by hypermail 2.2.0 : Sun Aug 25 2019 - 12:00:16 MDT