Re: Stressing in cluster??

From: Alex Rousskov (rousskov@measurement-factory.com)
Date: Tue Feb 06 2001 - 09:46:06 MST


On Tue, 6 Feb 2001, Mulyadi wrote:

> > Standard workloads are symmetrical and "uniform". That is, each
> > Polygraph client (server) runs the same number of equally configured
> > robot (server) agents. One can design a workload that is not
> > symmetrical and not uniform. Most configuration options are per-agent.
>
> I search the docs, but i can't find it. maybe you can give me some
> hint to do it??

How about looking at Robot and Server agent configurations in
workloads/simple.pg and workloads/include/polymix-3-guts.pg files?
Those files have good examples of agent configurations. Most agent
configuration options are documented at
        http://polygraph.ircache.net/doc/pgl/types.html
        http://polygraph.ircache.net/doc/pgl/types.html#Pgl:Robot
        http://polygraph.ircache.net/doc/pgl/types.html#Pgl:Server
 
> Mr. Rousskov, my final target is to maximizing paralelism on my
> cluster. If we run several scenario sequentially through single
> machine, i think it will to long . I think it will be more
> efficient if i can distribute several client and server pair on my
> cluster

Sorry, but I am still unsure what it is you want to do. What "cluster"
are you talking about? A cluster of Polygraph machines? A cluster of
caches? What is "scenario"? Please be specific. See if the answers
below help.
 
> 1. Can we configure Polybench so that we "fool" them so they act on
> homogenous machine?? I need this information, because my machines is not
> identical. In my lab, several machine are Pentium, the rest is Celeron.

You do not need to fool Polygraph -- Polygraph supports running a test
with several PCs as clients and servers. Simply describe your workload
in a PGL file and then start polyclt and polysrv processes on each PC,
specifying your configuration file as a parameter. See, for example,
        http://www.measurement-factory.com/docs/PolyMix-3/

BTW, Polygraph does not know or care what processor brand(s) you are
using.

> 2. Is there any tool to automate generating several scenario ??
> So, i run client server pair with cache hit eg 5% until 60%, with
> increment about 5% for example

There are several ways to test a cache with different hit ratios (if
that is what you are after, not sure). One way is to configure several
sequential PGL phases, one phase per hit ratio level you want to test.
Make each phase long enough to get steady state conditions. You may
also want to add "link" phases so that the hit ratio does not change
suddenly.

For example, PolyMix-3 has phases with ~5% hit ratio and
phases with ~55% hit ratio (see polymix-3-guts.pg). Phase object
configuration is described at
        http://polygraph.ircache.net/doc/pgl/types.html#Pgl:Phase
I do not think recurrence factors are documented there, but they are
the same as load factors, just applied to recurrence ratio as opposed
to load level. Again, see polymix-3-guts.pg. Keep in mind that factors
not specifically defined in a phase are inherited from the previous
phase.

There is no tool that will generate a PGL file based on a natural
language description, but you can write a simple C program or shell
script that produce Phase definitions you need.

The difficult part is to decide whether you want to change load levels
during each phase. If the request rate remains constant, you can
measure/compare the change in hit/miss response time between phases.
Unfortunately, Polygraph 2.5.4 cannot do the opposite (keeping
response time constant and auto-adjusting request rate instead). When
released, Polygraph 2.6 will probably have that feature.

Alex.



This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:17 MDT