Why does polygraph reduce offered load?

From: Konstantinos Nikoloudakis (nikolud@CS.UCLA.EDU)
Date: Thu Jan 11 2001 - 15:25:51 MST


Hello there,
I have been using polygraph to test a home grown cache on
a machine with lot's of RAM and fast CPU (but
not enough disk for such rate - the hit rate eventually settles to 30%).
 I'm using variants of the polymix-2 workload with version 2.2.9.
I have the following problem. After several hours of operation (> 2days)
at a load of 200 request/second (which is supposed to be
constant after the first 30min) polygraph reduces the offered rate
gradually down to ~150req/second (all the time req.rate and res.rate
go hand in hand). At the moment the decline starts,
the wait.level.mean graph shows rapid increasing trend
(and conn.*.level.mean also show increasing tendency, although
nothing excessive)
Throughout the run there are also a few errors (total of 0.02%)

I'm wondering whether the problem is because of a flaw in the
box that makes the performance deteriorate and forces polygraph
to react accordingly, or an experiment miscofiguration since
I have taken a couple of liberties with the configuration that might be
the cause of the problem. Namely:
1) use robot cloning on the same IP address rather than that
   configuring IP aliases
2) use fewer server agents than perscribed by the formula
   in http://polygraph.ircache.net/Workloads/PolyMix-2/
   (about 20 server agents, again on one IP but bound on different ports).

I know that taking such liberties does not produce the most accurate
testing environment, but had served its purpose to me so far for
shorter runs.
So if it's a configuration problem, why is it doing ok up to that
point and deteriorates later?

Thank you for your attention
Sincerely,
Kostas



This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:16 MDT