Re: [patch] sched: schedstat needs a diet

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Ken Chen (kenchen@google.com) wrote:
> On 10/18/07, Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> wrote:
> > Good question indeed. How large is this memory footprint exactly ? If it
> > is as small as you say, I suspect that the real issue could be that
> > these variable are accessed by the scheduler critical paths and
> > therefore trash the caches.
> 
> Maybe my wording was ambiguous, I meant to reduce cache line pollution
> when accessing these schedstat fields.
> 
> With unsigned long, on x86_64, schedstat consumes 288 bytes for each
> sched_domain and 128 bytes in struct rq.  On a extremely small system
> that has a couple of CPU sockets with one level of numa node, there
> will be 704 bytes per CPU for schedstat.  Given the sparseness of
> them, we are probably talking about 11-12 cache line eviction on
> several heavily used scheduler functions.  Reduce cache line pollution
> is the primary goal, actual memory consumption isn't really a concern.
> 

Generally speaking, if such cache trashing is an issue, why don't we
make sure that each task struct member is declared in this structure
following its access frequency ? (except for #ifdef blocks, which should
stay together) It could then statistically save a lot of cachelines.

Or is it already the case ? It doesn't look like it when I see:

struct list_head ptrace_list;

Just beside the 

struct mm_struct *mm, *active_mm;

pointers.

Mathieu

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/