Re: [PATCH 5/5] slub: Only IPI CPUs that have per cpu obj to flush

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Mon, 2011-09-26 at 11:35 +0300, Gilad Ben-Yossef wrote:
> Yes, the alloc in the flush_all path definitively needs to go. I
> wonder if just to resolve that allocating the mask per cpu and not in
> kmem_cache itself is not better - after all, all we need is a single
> mask per cpu when we wish to do a flush_all and no per cache. The
> memory overhead of that is slightly better. This doesn't cover the
> cahce bounce issue.
> 
> My thoughts regarding that were that since the flush_all() was a
> rather rare operation it is preferable to do some more
> work/interference here, if it allows us to avoid having to do more
> work in the hotter alloc/dealloc paths, especially since it allows us
> to have less IPIs that I figured are more intrusive then cacheline
> steals (are they?)
> 
> After all, for each CPU that actually needs to do a flush, we are
> making the flush a bit more expensive because of the cache bounce just
> before we send the IPI, but that IPI and further operations are an
> expensive operations anyway. For CPUs that don't need to do a flush, I
> replaced an IPI for a cacheline(s) steal. I figured it was still a
> good bargain

Hard to tell really, I've never really worked with these massive
machines, biggest I've got is 2 nodes and for that I think your
for_each_online_cpu() loop might indeed still be a win when compared to
extra accounting on the alloc/free paths.

The problem with a per-cpu cpumask is that you need to disable
preemption over the whole for_each_online_cpu() scan and that's not
really sane on very large machines as that can easily take a very long
time indeed.