MPAS partitioning 60-15 km mesh for 896 cores

[huangwei5934]

I have the 60 - 15 km mesh file:



-r--r--r-- 1 whuang users 29455387 Jul 14 2014 x4.535554.graph.info
-r--r--r-- 1 whuang users 2097256 Jul 14 2014 x4.535554.graph.info.part.1024
-rw-r--r-- 1 whuang users 1616205 Jul 21 2015 x4.535554.graph.info.part.112
-rw-r--r-- 1 whuang users 1681890 Jul 21 2015 x4.535554.graph.info.part.128
-rw-r--r-- 1 whuang users 1733122 Jul 21 2015 x4.535554.graph.info.part.144
-rw-r--r-- 1 whuang users 1774051 Jul 21 2015 x4.535554.graph.info.part.160
-rw-r--r-- 1 whuang users 1807601 Jul 21 2015 x4.535554.graph.info.part.176
-rw-r--r-- 1 whuang users 1835478 Jul 21 2015 x4.535554.graph.info.part.192
-rw-r--r-- 1 whuang users 1858977 Jul 21 2015 x4.535554.graph.info.part.208
-rw-r--r-- 1 whuang users 1879213 Jul 21 2015 x4.535554.graph.info.part.224
-rw-r--r-- 1 whuang users 1896621 Jul 21 2015 x4.535554.graph.info.part.240
-r--r--r-- 1 whuang users 1912214 Jul 14 2014 x4.535554.graph.info.part.256
-rw-r--r-- 1 whuang users 2027190 Mar 7 2015 x4.535554.graph.info.part.512
-r--r--r-- 1 whuang users 1523000 Jul 14 2014 x4.535554.graph.info.part.64
-rw-r--r-- 1 whuang users 1539697 Jul 21 2015 x4.535554.graph.info.part.80
-rw-r--r-- 1 whuang users 1550784 Jul 21 2015 x4.535554.graph.info.part.96
-r--r--r-- 1 whuang users 852601968 Jul 14 2014 x4.535554.grid.nc
-rw-r--r-- 1 whuang users 473809566 Jul 24 2015 x4.535554.tar.gz

But want to partition it for 896 cores, how do I run "parmetis"?
Usage: parmetis <graph-file> <op-type> <nparts> <adapth-factor> <ipc2redist> <dbglvl> <seed>

Thanks,


Wei

 

[mgduda]

I've never actually used 'parmetis', but rather, the 'gpmetis' program. With gpmetis, a command like the following should work:

gpmetis -minconn -contig -niter=200 x4.535554.graph.info 896

There were some instances where Metis would produce non-contiguous partitions, so the '-contig' option seems like a good idea. Using 200 iterations is something of an arbitrary choice, and I don't have any quantitative results to suggest that this is better than the default number of iterations, but given that gpmetis runs so quickly anyway, '-niter=200' doesn't seem to hurt.