Erik de Castro Lopo wrote:
Multiprocessor Opteron systems do implement NUMA (each CPU has a memory controller with banks attached), but I don't think this is the source of the slow down - the HyperTransport links between CPUs and memory are very, very fast (and specifically designed to avoid these sorts of issues).- Is there any chance your machine is NUMA (Non Uniform Memory Architecture)? This basically means that all CPU can access all the memory in the machine, but that for each processor some memory is local and fast while other memory is furtheraway (ie on another card on the other side of some shared bus) and hence slower? If so, you may find that the pointwhere you get your slowdown coincides with filling up local memory and starting on more remote memory.
I think Matt might be on the right track with swap, but I'm at a loss to figure out how to get around it / pin down the bug.
- Matt _______________________________________________ coders mailing list [email protected] http://lists.slug.org.au/listinfo/coders
