3 ms·
Intel did some work on this for the Xeon Phi that had fast on-package memory and slower main memory. IIRC they use fake NUMA nodes to separate the RAM.
by wmf 8y ago
Intel did some work on this for the Xeon Phi that had fast on-package memory and slower main memory. IIRC they use fake NUMA nodes to separate the RAM.
- toast0 8y ago> fake NUMA nodes Is it really fake if the memory access is non-uniform? Maybe it's unconventional?
- wmf 8y agoYou got me there. I think conventionally each cpu is a child of one NUMA node so in this case they had to make a NUMA node with no cpus, which isn't quite correct. https://colfaxresearch.com/knl-mcdram/#fig-3 https://colfaxresearch.com/knl-mcdram/#fig-3