RE: [PATCH] Binding processes to selected CPUs
avik@talmai.com
Wed, 13 Oct 1999 11:28:07 +0200
>From: Matti Aarnio [mailto:matti.aarnio@sonera.fi]
>...
>> I redid my tests, this time against physical memory,
>> and they show that there is no performance gain with
>> 128kB caches. Perhaps machines with more cpus and
>> larger caches will benefit, but the smaller ones will
>> not.
>
> People here seem to presume that all the world SMPs are
>UMA systems.
>
> Taxonomy of parallel computers (by David Black, back then at CMU):
> - UMA Uniform Memory Access (your usual tightly coupled SMP)
> - NUMA Non-Uniform Memory Access (processors have good access
> to memory which is local to them, but slower access to
> memory at other CPU/MEM boards; beasts like SUN E 10000)
> - NORMA NO Remote Memory Access (your average Beowulf)
> Spice this with Cache Coherence, and you get CC-NUMA (and CC-UMA).
>
> (That was after "In Search of Clusters" by Gregory F. Pfister)
>
> Your observations may well hold at UMA systems.
>
> Binding facilities make a *lot* sense at NUMA systems.
> (Also memory allocation policies must be tuned
>accordingly; allocing
> process local memory from remote board hurts process, AND system
> performance when processor most go thru system memory crossbar to
> access that memory causing contention for that resource..)
>
All true, but I lack the facilities to measure the performance
of my patch on a NUMA :(
Does linux support a NUMA machine?
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/