
[Science] [BAS home] [Met home] [Beowulf home]	Antarctic Meteorology

Running the model on a sun

The sun is a

SunOS bsupsdb 5.8 Generic_108528-15 sun4u sparc SUNW,Sun-Blade-1000

according to uname -a. Our people say it is an ultrasparc 3, and really a sunblade 2000 at 980 Mhz, no matter what the above may say...

How

I built the run under linux as normal, tarred up the source (using "follow symlinks of course), untarred under sun, compiled, found the dump files etc (the start file built under linux needs byteswapping, the ancils from the umsystem.tar don't), altered various paths in the enviroment, and hey preseto, it ran. But slowly.

Timing

At 64-bit, compiling with:

f90 -e -c -xarch=v9  -xtypemap=real:64,double:64,integer:64

it runs at a feeble 15.4 mins/model_day. Using "cpu" I see it running at 96-ish%, which implies (I think) that it isn't waiting for disk/memory much. I admit to running it from a slow disk system, but that shouldn't matter (except it compiles like a dog).

Adding the -O4 option makes a considerable difference, down to 6.4 mins/day. This is 50% faster than one athlon 1600, or 50% slower than two... however, the cost of the machine is about GBP 7,600 = VAT, which is about 4* the cost of a dual-athlon node.

See timings for comparison with other machines.

Try adding the -fast flag... various routines won't compile properly, e.g. fill3a says:

fill3a.f
cg: assertion failed in file ../src/codegen/leafgen.cc at line 2617
cg: gen_addrof_register: bad addrof
cg: 1 errors

This turns out to be caused by the -xvector=yes component of -fast. Further tests show the stupidity of the system: -fast includes -xtarget=native, which causes "wrong elfclass" messages, because it it too stupid to set -xtarget=native64 as necessary.

Compiling with an explicitly set subset of the -fast options fails at link time... still investigating.

Run-time errors

The run died with:

 ******  FORTRAN RUN-TIME SYSTEM  ******
 Error 1141:  status is specified and not 'OLD' for connected unit
 Location:  the OPEN statement at line 4790 of "meanctl1.f"
 Unit:  8
 File:  ./tmp/yaboa.pipe
 Status:  REPLACE

which is at least a big improvement over Fujitsu as to explicitness of error message.

For the second run I deleted all references to unit 8 wherever I found them - its only the talking to the server, which we don't use. Then there is no problem. ~

Past last modified: 13/9/2002 / wmc@bas.ac.uk
© Copyright Natural Environment Research Council - British Antarctic Survey 2001