On Fri Apr 17 14:21:03 EDT 2009, tlaro...@polynum.com wrote: > On Fri, Apr 17, 2009 at 01:29:09PM -0400, erik quanstrom wrote: > > > In some sense, logically (but not efficiently: read the caveats in the > > > Plan9 papers; a processor is nothing without tightly coupled memory, so > > > memory is not a remote pool sharable---Mach!), > > > > if you look closely enough, this kind of breaks down. numa > > machines are pretty popular these days (opteron, intel qpi-based > > processors). it's possible with a modest loss of performance to > > share memory across processors and not worry about it. > > NUMA are, from my point of view, "tightly" connected. > > By loosely, I mean a memory accessed by non dedicated processor > hardware means (if this makes sense). Moving data from different > memories via some IP based protocol or worse. But all in all, > finally a copy is put in the tightly connected memory, whether huge > caches, or dedicated main memory.
why do you care what gives you the illusion of a large, flat address space? that is, what is special about having a quick path network instead of, say, infiniband or ethernet? why does networking imply ip networking? my point is that i think we need to recognize that there vast differences in performance between, say, local memory, memory across the quickpath bus, memory on the the next machine, and these differences may vary greatly between one set of machines and another. then, the 64ยข question is, how does one use this to one's advantage without assuming ahead of time what's faster than what. (one could easily imagine a 40gbps ethernet connection being competitive with a 3-hop numa connection.) - erik
