Summary of Cray

My key points from (mostly from Paul Selwood's report):

  • The compute is slightly faster for Cray than IBM
  • Cray has much lower bandwidth than IBM. This particlarly restricts speed of
    • UM dynamical core
    • NEMO
    • And it seems: GLOMAP-mode
  • Hyperthreading is useful for a small number of nodes; not for lots of nodes
  • Unlike IBM, seems better to have more MPI tasks in North-South than East-West
  • Variation on timing for a GC3 run is 11-12% (according to JC, probably 10 days)