Re: [PATCH 00/14] Zoned VM counters V5

[Martin Bligh <mbligh@google.com>]

Christoph Lameter wrote:
> On Wed, 21 Jun 2006, Martin J. Bligh wrote:
> 
> 
>>Having the per-cpu counters with a global overflow seems like a really
>>nice way to do counters to me - is it worth doing this as a more
>>generalized counter type so that others could use it?
> 
> 
> Yes later patches also use the counters for other things. Please check out 
> the patch that uses these for numa counters etc.

OK, but looked like the original implementation was sort of tied to
zones / the VM, at least in terminology, and code placement. I'll look
at it again ...

>>OTOH, I'm unsure why we're only using 8 bits in struct zone, which isn't
>>size critical. Is it just so you can pack vast numbers of different stats into
>>a single cacheline?
> 
> 
> I would like to add some stats in the future. 8 bits is sufficient if the 
> threshold is less than 64 (currently its 32). If we ever get higher then 
> we can simply go to a bigger base size.
> 
> However, the space used by that array is
>  
> <nr-of-counters>*<nr_of_processors>*<nr_of_zones>
> 
> There are systems that have around 1k nodes and 4k processors. Lets say 
> we have 16 counters then we get to
> 
> 1k*4k*16 = 64Mbyte just for the counters.
> 
> This doubles for a short and quadruples for an int.
> 
> Also smaller counters help keep the pcp structure in one cacheline and 
> reduces the cache footprint. 

Sure, but for a normal sized system, the smaller the per-cpu portion,
the more atomic ops you'll end up doing, surely?

M.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>