--- John Hawkes wrote
> From: "Juergen Doelle" <JDOELLE@...>
> > My impression is that additional workload from a measuring tool like
> > lockmeter changes the system behavior, so that the results with
> > lockmeter running do not reflect the situation when lockmeter is
> > not running.
> Yes, this is quite likely. Even when a lockmeter-capable kernel has
> lockmeter turned off, the spinlock/rwlock calls are still procedure
> calls instead of inline asm. I would hope you're doing the basic
> benchmarking using a non-lockmeter-patched kernel, and only using
> lockmeter-capable and kernprof-capable kernels to peer into kernel
> behavior to understand the basic benchmark results. (I don't recommend
> using a kernel with both the lockmeter and kernprof patches, as the
> totality of system perturbation is just too large.)
> John Hawkes
Of course, a kernel with lockmeter patch was only used for the
lockmeter measurements. Benchmarking was done with a plain
vanilla 2.4.4 kernel and the described lock patches.
To make it clearer:
My impression is that in this particular case the additional workload/
overhead from lockmeter has changed the system behavior. The lock
behaviour with lockmeter running seem not to reflect the situation when
lockmeter is not installed.
The cause for that impression is that lockmeter reports a great
one CPU was freed from spinning, but the throughput values (from a kernel
without lockmeter patch) does only increase slighty.