Messages in this thread | | | Date | Tue, 28 Apr 2009 09:40:31 +0200 | From | Andi Kleen <> | Subject | Re: [PATCH 5/5] proc: export more page flags in /proc/kpageflags |
| |
> I think i have to NAK this kind of ad-hoc instrumentation of kernel > internals and statistics until we clear up why such instrumentation
I think because it has zero fast path overhead and can be used any time without enabling anything special.
> measures are being accepted into the MM while other, more dynamic
While the dynamic instrumentation you're proposing has non zero fast path overhead, especially if you consider the CPU time needed for the backend computation in user space too.
And it requires explicit tracing first and some backend that counts the events and maintains a shadow data structure covering all of mem_map again.
So it's clear your alternative will be much more costly, plus have additional drawbacks (needs enabling first, cannot take a snapshot at arbitary time)
Also dynamic tracing tends to have trouble with full memory observation. I experimented with systemtap tracing for my memory usage paper I did a couple of years ago, but ended up with integrated counters (similar to those) because it was impossible to do proper accounting for the pages set up in early boot with the standard tracers.
I suspect both have their uses (that's indeed some things that can only be done with dynamic tracing), but they're clearly complementary and the static facility seems useful enough on its own.
I think Fengguang is demonstrating that clearly by the great improvements he's doing for readahead which are enabled by these patches.
-Andi
| |