Re: Ideal cache performance

From: Alex Rousskov (rousskov@ircache.net)
Date: Thu Jul 27 2000 - 09:34:11 MDT


On Thu, 27 Jul 2000, Alan Smithee wrote:

> An ideal cache running PolyMix-3 would have a 55% hit ratio, right?

An ideal cache running PolyMix-3 would have a measured hit ratio
matching the offered hit ratio. The "ideal" level is hard, if not
impossible, to calculate because it is not clear (to me) what a "hit" is
when you have IMS requests. You sort of have different classes of hits
now.

Polygraph currently walks around the problem by considering only "basic"
transactions when reporting offered and measured hit ratios. We may end
up comparing client and server logs to provide with more detailed
statistics in the future.

> What would be the percentage of requests for which an ideal cache:
> - refreshes an expired object in its cache by requesting an IMS from
> the origin server
> - generates a 200 response to an IMS request
> - generates a 304 response to an IMS request

I cannot answer this question for at least two reasons:

        - I do not think there is a [single] ideal behavior with
          respect to handling IMS requests; there are several
          tradeoffs that caches may resolve in different ways
        
        - I do not have an implementation of an "ideal" cache to
          measure one possible distribution, and I lack brains
          and other resources to build an analytical model for
          such a complex workload as PolyMix-3

While it is unfortunate that we do not know "ideal" ways to resolve the
tradeoffs, it is the price for making the workloads more realistic.
Adding IMS and other types of requests to PolyMix was an important
improvement, demanded by many users. The result is a better, but harder
to analyze, workload.

Please note that the situation is not unique to PolyMix-3. For example,
if you ask me how many requests per connection and ideal cache would
serve under PolyMix-2, my answer would be the same. Moreover, even
achieving ideal hit ratio may not be the best strategy as many vendors
have demonstrated in the past: the (hit ratio, throughput) tuple can be
"optimized" in different ways...

Polypxy that comes with Polygraph cannot handle IMS requests yet. When
that functionality is added, you will have one more "reference point".
Meanwhile, you may try to use your competitors' caches to get an idea of
how they resolve the same tradeoffs.

$0.00,

Alex.



This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:14 MDT