29 Dec
2017
29 Dec
'17
3:27 a.m.
Paul Davis wrote:
On Thu, Dec 28, 2017 at 2:11 PM, Jeremy Henty
<onepoint(a)starurchin.org>
wrote:
Until I read the Wikipedia NUMA page just now I didn't realise that
NUMA involved multiple processors having their own caches of the same
data. Thanks for bringing me up to speed.
Regards,
Jeremy Henty
Will Godfrey wrote:
AFAIK, NUMA is dead for everything except a few research systems.
Parallel/multi-processor systems these days are all "symmetric" (all
processors have symmetrical access to all memory).
NUMA is really, really, really hard to get right. Why? Cache
invalidation. Several companies, organizations, etc. have
tried. Last time I looked (and it has been a while, but I was quite
involved with this stuff in the mid 1990s), everybody failed. If I've understood that correctly you can
also ensure that they are
also on the same socket, which apparently improves memory access.
I think this is what is meant by NUMA (Non-Uniform Memory access).