1999-01-05 09:38:12

by Albert D. Cahalan

[permalink] [raw]
Subject: Re: Good MM benchmarks utilities?

Steve Bergman writes:

> I've been doing some benchmarking of the 2.2pre kernels and various
> patches in development. I am looking for memory management benchmark
> utilities. Something like Bonnie for vm subsystem benchmarking.
> I'm concerned about the comprehensiveness of my own simple tests.
> What I'm finding is that while 2.2 is outperforming 2.0.36 in some
> situations, it is falling far behind in others. Particularly the very
> low memory case (6M with X running) 2.0.36 performs far better (Even
> though the overall 'available memory' as reported by top is actually
> slightly *greater* for 2.2. Andrea Arcangeli's recent patches have
> improved things a great deal but I still haven't found anything to
> beat 2.0.36 in very low memory. It would help to have a suite that
> gives a more comprehensive view of mm performance.
> Also, no matter how good the suite is, it will only tell me about the
> performance of my particular hardware layout. The more people who
> test and submit results, the better 2.2/2.3 mm will be.

If gcc could be IO-bound, you'd have a great benchmark.
You could get a 450 MHZ Xeon or 666 MHz Alpha with old IDE drives,
or slow down your IO in software to adjust for a normal CPU.

There might be a place in the block device code where you could add
a nice delay. You could get the NFS swap patches and use a PPP link
over a 14.4 modem. I suspect that would make gcc highly IO-bound.

If you had the kernel log user IO operations and page faults, you
could have a benchmark play back gcc IO behavior without all the
normal CPU usage.