Re: [RFC] Scalable statistics counters using kmalloc_percpu

Ravikiran G Thirumalai (kiran@in.ibm.com)
Tue, 30 Jul 2002 16:55:48 +0530


On Mon, Jul 29, 2002 at 11:23:06AM -0700, Andrew Morton wrote:
> ...
> Well Rusty's point about just using the percpu API seemed
> reasonable - that's basically equivalent to statically defining
> the data. In what situation is dynamic allocation needed?
>

As Dipankar pointed out, counters part of dynamically allocated structures;
Like:

1. k_atm_dev_stats in atm_dev -- they are all atomic_t types right now.
we must see a major benefit with statctrs here.
I don't have access to atm infrastructure to test it though :(((
2. Atomic counters like dst->__refcnt; we might be able to hack
route cache so that the periodic garbage collector can tolerate
stale reads on the refcnt. A dst_entry might get gc'd late, nevertheless
it would boost performance on networking loads. (I still have to
investigate if we can do it at all...).

This is what I have in mind for now. Of course, if these primitives
are available to all, programmers will have more flexibility in
designing datastructures. We can then afford to have non atomic_t
global counters for debugging etc., anywhere in the kernel; modules/dynamic
structures...

Thanks,
Kiran
-
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/