[MITgcm-support] Parallel performance

[Dimitris Menemenlis]

Stefano and Kevin, "if" nfs I/O is the issue try "useSingleCpuIO=.TRUE." 
and make sure that you start the model from an I/O node, i.e., a node 
that is connected directly to a disk.  This will read and write 
everything from the root CPU.  For the platforms that I am familiar 
with, i.e., SGI Origin and Altix, usesinglecpuio is faster than reading 
and writing to different files from each CPU and the globalfiles option 
either does not work (has drop outs) or is excrutiatingly slow.  On a 
platform with a truly distributed I/O set-up, I expect 
"useSingleCpuIO=.TRUE." to be slightly slower than reading and writing 
to separate files from each processor, but only slightly so and it is 
infinitely more convenient as it avoids the step of copying input files 
to all the disks and of joining the output from multiple files 
afterwards.  Dimitris

-- 
Dimitris Menemenlis <menemenlis at jpl.nasa.gov>
Jet Propulsion Lab, California Institute of Technology
MS 300-323, 4800 Oak Grove Dr, Pasadena CA 91109-8099
tel: 818-354-1656;  fax: 818-393-6720