Parallel NCO Benchmarks

• The benchmarks below are an interim measure of the project to parallelize the NCO's. This has been done in 2 ways: 1) using OpenMP threading and 2) using MPI (Message Passing Interface). The first is appropriate for shared memory architetures where a large number of processors share a large common pool of memory. The latter is appropriate for clusters of (slightly) heterogeneous workstations as well as Shared memory machines.
• Currently the OpenMP has been tested out to 8 threads on an 8 CPU machine and the MPI version has been tested to 4 processes on a 2 CPU machine. We will be testing more configurations in the near future.
• the ncap operator has not yet been parallelized for MPI and so the MPI benchmarks are somewhat irrelevant
• The MPI processes listed are the number of 'MPI workers'; there is an additional master process that only coordinates workers
• The plots that contain more than about 6 points contain data from previous versions of nco; the data from the most recent version tends to be the fastest. Newer versions of these benchmarks will be 'version-pure'.
• [Oct 28 05]The plots are being converted to a normalized format so that scaling among NCO's is more directly comparable. This is shown in the AIX benchmarks below and the others will be converted and brought up to date soon.

OS & Hardware	Compiler	OMP/MPI	NCO Ver	Benchmark
ESMF,AIX 5.1, dedicated 8 Power4 CPUs, 16GB RAM direct attached disk (/scratch) NORMALIZED	IBM xlc_r v6	OMP	20051027 / 20051022	click thumbnail for full PDF benchmarks
ESMF,AIX 5.1, dedicated 8 Power4 CPUs, 16GB RAM direct attached disk (/scratch) NORMALIZED	IBM xlc_r v6	OMP	combined	click thumbnail for full PDF benchmarks
clay-local: Linux 2.6.11 SMP-AMD64, 2x@GHz, 4GB RAM local ext3 filesystem NOT NORMALIZED	mpicc (gcc 4.0.0)	MPI (mpich2)	20050928 / 2005/09/15	click thumbnail for full PDF benchmarks
clay-OpenDAP: Linux 2.6.11 SMP-AMD64, 2x@GHz, 4GB RAM Connecting to OpenDAP server (similar configuration) 1 hop away over 1Gbs ethernet. NOT NORMALIZED	mpicc (gcc 4.0.0)	MPI (mpich2)	20050928 / 2005/09/15	click thumbnail for full PDF benchmarks

More benchmarks coming...