TalkChess.com

[Daniel Shawul]

I will do that I guess. He is currently doing a postdoc and I didn't want to disturb him. The best I can do for a guy who found a way out of this CCC grave
Performance guidelines for AMD opteron ccNuma pretty much says I make sure local memory is allocated for each thread.
"Don't do explicity memory allocation, os does it better"
"Don't set thread affinity, os does it better" etc
So in the end the programmer doesn't need to do much to get optimal performance other than "first touch" rule, and maybe some cache optimization specific for numa.
Anyway the scale of SMP zappa used (512 cores) and the itanium architecture are interseting to know the limits of SMP, and when to decide to go message passing instead. Surely there will be a lot of inefficiency for but how much. Most (lazy people) here actually use mpi to program even for numa machines but that probably doesn't translate to chess as efficiency is we care about.
_________________
https://sites.google.com/site/dshawul/
https://github.com/dshawul