Evaluation of parallel program execution with Scalasca profiler
Rőczei Gábor
<>
NIIF Intézet
The supercomputers contain more and more processor cores due to the continuous development of the supercomputer manufacturers. For example, the world fastest supercomputer was the Tianhe-2 (Milkyway-2) in November 2013. This data is from the top500 list. It consists of 3120000 processor cores. This shows how complex systems are involved. We will only be happy if we are able to use these processor cores in accordance with supercomputer softwares and they can run fast. Is not the same ones that a scientific calculation ("in silico" experiment) runs for a month, or just a few days. This is especially interesting if we would like to run the same calculation as many times, but with different parameters. These software solutions can exploit parallelism using a parallel paradigm. The two most common parallel paradigms are the MPI and OpenMP.
The profiler is a program intended to evaluate an other program, how it is run well or poorly. In case of multi-processor systems Scalasca profiler is a perfect choice for this purpose. This is a free available profiler program, which can provide support for parallel programs (MPI, OpenMP and by hybrid case (MPI + OpenMP) as well) to optimize performance. From this page we can download it: http://www.scalasca.org and it is available for free download under the New BSD open-source license.
I will demonstrate the main features of Scalasca profiler with some simple examples during the presentation. I really want to accomplish that it become a useful tool for the supercomputer users in Hungary.