On 29-set-09, at 19:08, Samuel Thibault wrote:
Fawzi Mohamed, le Tue 29 Sep 2009 18:55:27 +0200, a écrit :
cpuset -> cpuset_ptr (or just a flag that says if the structure has
it, and thus two structures, a long one with it and a short one
without, differing only in the tail if you really want to be hacky).
Then cpuset is generated on the fly for the deepest level (like less
than 4-8 proc -> lots of memory savings on large machines).
(cost 1 function, and copying or building the cpuset)
Even for a machine with 1024 processors that amounts to just 128KB
saving. Even if you replicate it on each NUMA node, such nodes
usually
have GBs of memory. Is it really worth spending efforts to save
it? :)
ok I will stop asking, those were just things that I was thinking when
I began to work with the code... because my impression is that the
memory per core is not increasing much anymore and some processor
design advocate massive multithreading to hide memory latency, but
maybe memory starved blue gene/L are indeed of the past, I just
checked and the /Q version should have 1GB per core from what they
say...
I did just throw them there, make what you want of them...
You know the newcomers worry about the investment and choices they are
doing :)
ciao
Fawzi
