hui_wan@pnnl_gov
New Member
Hi,I did a series of single-instance and multi-instance runs on Titan at Oak Ridge. For the same kind of simulations, the multi-instance runs cost significantly more core hours. Here are some numbers from the cesm_timing.xxxxx files: ------------------------------------------------------------------------------- Single-instance Multi-instance (1x2000 cores) (32x1888 cores) Wallclock WallclockDRIVER_INIT 61.501438 344.224579 (Called = 2)DRIVER_RUN_LOOP 1219.016846 1695.854736DRIVER_CPL_RUN 24.72876 134.189667DRIVER_CPL_COMM 27.324247 529.453613--------------------------------------------------------------------------------Compset: FC5PE layout: same number of processes for all active components; same number of instances (1 or 32) for atm, lnd, rof, ice, ocn; one instance for cpl, one instance for glc and wav which are stub components.Code: ACME code which branched from CESM at 1_3_beta10. I'd like to ask whether similar behavior has been seen with CESM, and what I could do to improve the performance. Thanks in advance.Hui