Re: [patch 3/4] net: Percpufy frequently used variables -- proto.sockets_allocated

[Andrew Morton <akpm@osdl.org>]

Eric Dumazet <dada1@cosmosbay.com> wrote:
>
> Ravikiran G Thirumalai a écrit :
> > On Fri, Jan 27, 2006 at 12:16:02PM -0800, Andrew Morton wrote:
> >> Ravikiran G Thirumalai <kiran@scalex86.org> wrote:
> >>> which can be assumed as not frequent.  
> >>> At sk_stream_mem_schedule(), read_sockets_allocated() is invoked only 
> >>> certain conditions, under memory pressure -- on a large CPU count machine, 
> >>> you'd have large memory, and I don't think read_sockets_allocated would get 
> >>> called often.  It did not atleast on our 8cpu/16G box.  So this should be OK 
> >>> I think.
> >> That being said, the percpu_counters aren't a terribly successful concept
> >> and probably do need a revisit due to the high inaccuracy at high CPU
> >> counts.  It might be better to do some generic version of vm_acct_memory()
> >> instead.
> > 
> > AFAICS vm_acct_memory is no better.  The deviation on large cpu counts is the 
> > same as percpu_counters -- (NR_CPUS * NR_CPUS * 2) ...
> 
> Ah... yes you are right, I read min(16, NR_CPUS*2)

So did I ;)

> I wonder if it is not a typo... I mean, I understand the more cpus you have, 
> the less updates on central atomic_t is desirable, but a quadratic offset 
> seems too much...

I'm not sure whether it was a mistake or if I intended it and didn't do the
sums on accuracy :(

An advantage of retaining a spinlock in percpu_counter is that if accuracy
is needed at a low rate (say, /proc reading) we can take the lock and then
go spill each CPU's local count into the main one.  It would need to be a
very low rate though.   Or we make the cpu-local counters atomic too.

Certainly it's sensible to delegate the tuning to the creator of the
percpu_counter, but it'll be a difficult thing to get right.