Re: [PATCH UPDATED] percpu: use dynamic percpu allocator as thedefault percpu allocator

From: Christoph Lameter
Date: Wed Apr 01 2009 - 22:26:41 EST


On Thu, 2 Apr 2009, Ingo Molnar wrote:

> Note that my argument was different though: that assumptions about
> variable correlation are very hard to track and validate, and that
> IMHO we should be using __read_mostly generously (we know _that_
> attribute with a rather high likelyhood), and we should group the
> remaining variables together, starting at a cacheline aligned
> address.

But then you decrease the density of accessed to the __read_mostly
sections. The cachelines are not hot in the caches anymore which is an
average performance reduction.

> A sub-sub argument was that perhaps we should not split .data and
> .bss variables into separate sections - it doubles the chance of
> false cacheline sharing and spreads the cacheline footprint.

False cacheline sharing is something normal that comes with the cpu
caching schemes. As long as there is no significant impact on performance
we are fine with it. Extensive measures to avoid false cacheline sharing
on unimportant variables increases the cache footprint of code.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/