[Mitgcm-support] RE: timings

mitgcm-support at dev.mitgcm.org mitgcm-support at dev.mitgcm.org
Wed Jul 9 15:33:46 EDT 2003


Chris,

Blocking_exchanges timings:
   6.9ms +-0.1ms on 1 proc
  24.6ms +-0.3ms on 6 procs with yesterdays/your version of exch_rx_cube
               (code is overall slower than 1 proc)
  19.1ms +-0.1ms on 6 procs with all sends then all recvs
               (code is marginally faster than 1 proc)
  19.2ms +-0.1ms, as last but without MPI_Wait

Solve_for_pressure timings:
  0.189s +-0.001s on 1 proc
  0.230s +-0.010s on 6 with yesterdays code
  0.178s +-0.005s on 6 with todays (all send then recv)

This is all with the same exch_uv_rx_cube

BTW, I noticed that sometimes the usertime > wall clock (single and
multi-proc). This shouldn't happen! Are the timers that inaccurate?

A.



> -----Original Message-----
> From: Chris Hill [mailto:cnh at plume.mit.edu]
> Sent: Thursday, May 22, 2003 9:35 AM
> To: adcroft at ocean.mit.edu
> Subject: timings
> 
> 
> Did you make an estimate for the 6-way exchanges timing?
> The CS + atmos is around 0.11 secs per time step uniprocessor 
> - so if the 6-way comms is near that we have a problem!
> 
> Chris
> 




More information about the MITgcm-support mailing list