Using the komp cluster

Instructions for using the komp cluster

24 dual-400Mhz-processor, called komp01-komp24. komp01 is the "front end" with about 400 Gb of disk space, and should only be used for compiling, visualization etc. Contains data disks d1-d6 and RAID r1.
1 dual-850 MHz, called dual850, with data disks e1-e4.
1 single Athlon 1333, called niban
1 single Athlon XP1900+, called sanban
1 single Pentium4 2GHz, called yonban

From the outside world you can only log in to komp01, the 'server' node. However you can then rlogin to other komps, and your home directory and data disks are the same (thanks to automount).
To see how busy the cluster is, type 'cl_load'
MPI is the communications library used. Since there is no queuing system, it is necessary to manually specify which nodes to run on, after first checking which ones are free!
There are 2 CPUs per komp and hence 2 MPI processes can be run per komp.

For scalar (single-node) programs without MPI use either the GNU compilers g77, gcc, g++, Portland Group compilers pgf77, pgf90, pgcc, pgCC. or the Intel compiler ifc. The Portland Group compilers generally produce code which is faster by ~5-20% for single precision or 50-100% for double precision. The PG compilers also have the possibility of auto-parallelizing for the 2 processors in each komp (SMP) although we haven't tested how well this works.
For compiling parallel programs use mpif77, mpicc or mpiCC. These are presently coupled to the GNU compilers g77, gcc and g++. It would probably be better to couple them to the PG compilers but I haven't had time to figure this out yet.

Avoid running jobs on komp01 because its CPUs are often busy with I/O, tape backups, compiling, etc.
If someone is already running on a particular komp, don't start another job on it.
No more than 1 job per CPU (2 jobs per komp).
Don't fill up the disks!