NEMO3D

This is a summary of the performance evaluation of NEMO3D. Our initial focus was on finding hot spots in the code where most of the computational work is being done. In all cases the NEMO3D benchmark_lanc_thin_no_strain (262144) with recomputed matrices was used. Overhead was calculated on 16 processors.

Instrumentation overhead

Run Type	Runtime (seconds)	Overhead %
Uninstrumented runtime	372	NA
Routine+loops instrumented	392	5.4

Significant Loops

We have found two loops in the source file "h_cvectr_multi.c" that together account for about 90% (with 16 processors) of the runtime of the NEMO3D application. Loop 1 starts at 1235 and ends at 1841. Loop 2 starts at 1270 and ends at 1760.

Runtime Breakdown Charts

These charts show the runtime breakdown of NEMO3D on processor counts 16,32,64,128,256. Each experiment was run on PSC's SGI Altix system (pople).

Legend

Entire Experiment Loop 1 Loop 2