Today I met a famous FreeBSD developer, David O’Brien (a.k.a. obrien@) and he gave us a presentation about the AMD64 architecture, which cleared my misunderstanding about the AMD64 architecture set by some marketing guys presenting in Tsinghua University.

The MP model of AMD64 is not simple SMP, but NUMA. In other words, the integration of memory controller DOES increase the performance in MP case and provides better scalablity. Of course this needs the operating system to be NUMA-aware (Linux does, but FreeBSD’s work is ongoing), but the NUMA architecture has better scalablity because it eliminates the bottleneck created by northbridge chips (all chips shares one northbridge chip and hence it’s the bottleneck), as every CPU controls its own memory and when they are going to accept foreign memory, they access through other CPUs (this in some views is expensive, but could be solved by software, especially scheduling, etc).