FEATURES & COMMENTARY
On June 30, 2000 there was a Feature & Commentary “PERFORMANCE MEASUREMENT ISSUES IN HPC.” It was reported about the search for a realisic metric to measure the performance.
An important aspect of HPC is Scientific Computing. There we have an easy and realistic metric: the vector triad
a(i) = b(i) + c(i) * d(i).
This operation delivers supervector or superscalar speed under most general conditions. It is the basic operation of engineering applications. The performance of this operation does not depend on the processor speed (like Linpack). It depends on the cache bandwidth for data from cache and from the memory bandwidth for data from memory. As usual application software has no cache blocking algorithms, the performance for data from memory is my metric and this scales well with many engineering applications.
In my book “Scientific Supercomputing” and in the ADDENDUM to the book measurements of many actual processors are presented and the effect of the bandwidth is demonstrated, see URL
The discussion of the measurements reveals the weak points of the measured processors.
Rechenzentrum der Universitaet Karlsruhe
Email: [email protected]