Why Distribute?

• CC-NUMA architecture optimization. If:

  • all data initialized on one node
    • memory bottle neck
  • all shared data on one node
    • serialization of the CC
  • data used from different node
    • large interconnect traffic
    • average net latency on SGI3800:

16 proc

memory

Processor(s)

~(1/2 log2(p/16)+1)*50 ns

Previous slide

Next slide

Back to first slide

View graphic version