[MITgcm-support] IO on NEC SX machines
Martin Losch
Martin.Losch at awi.de
Fri Nov 16 06:33:44 EST 2007
Hi there,
in my pathological configuration I have daily forcing with a 12h time
step (for a spinup), so every other time step new forcing fields are
read. This produces an overhead on any machine (I guess), but on our
SX8R, I find this overhead to be very large: apparently in some of my
runs 2/3rds of the time is spend in reading the forcing. I don't
observe any significant differences between runs with and with
useSingleCPUio = .true., one or more CPUs. I ran exactly the same
configuration on our Cray XD1 on 18cpus and there is hardly any
overhead (< 5%, regardless of useSingleCPUio=.true. or not).
I have talked to Jens-Olaf about this (maybe he can correct me, if I
am mixing things up here), and it appears that the chunks that are
read in mdsioreadfield are too small to take advantage of the
otherwise fast GFS filesystem of the machine. This is my domain size
(and forcing field size): 180x120 grid points and my forcing is in
double precision so one record is 172800 bytes (correct?).
The obvious solution is to read forcing in larger chunks (e.g. per
year) and keep it in memory, but that would require considerable
recoding.
Is there any experience with this problem on other NEC SX machines,
GFS file systems? Are there some hidden switches in the MITgcm that
would help to improve this IO performance?
Martin
More information about the MITgcm-support
mailing list