Analyzing for Parallelism (cont'd)

• Where are the data?
  • How to maximize computation/communication ratio?
    • computation rates (Cray T3D):
      • From peak 150 MFLOPS 1 FLOP= sec

      • From typical 20 MFLOPS 1 FLOP= sec

    • communication rates (Cray T3D)
      • Peak comm. latency (shmem_put)
        1 Xfer = sec

      • Peak comm. B/W (shmem_put) 120MB/s
        1 Word = sec

    • Choice of domain decomposition important
      • Surface-to-volume ratio of subdomains

Generated with CERN WebMaker