some MIRIAD benchmarks notes
Notes on parallel machines. Example: a 24 node 1GHz (AMD) cluster, built
in May 2001.
- although a single CPU still runs around 3 MirStones, the
I/O is bound by a SCSI-RAID, mounted on the server, and NFS
will slow it down to a MirStone of 2.1 !!!
- running two mir.bench's in parallel give about 1.7, and running
4 in parallel to about 1.3
- it is mostly INVERT that causes mir.bench to slow down.
- if TMPDIR is set wrong, it can cause NFS traffic for scratch files,
and severely slow down the benchmark
- on linux the Intel 5.0 compiler runs about 30% faster than
the gnu (2.95.3) compiler.