Section: Application Domains

Signal processing

ParSONS is a parallel sound sources classifier developed at SUPELEC by Stephane Rossignol. It allows to identify different kinds of sound sources in an audio signal file (like a radio or TV archive file), and to tag different part of the file. Then it is possible to know which part of the archive file include human voice, or music, or more accurately jazz or rock music. From an algorithmic point of view, this application performs a lot of Fourrier Transform computations and a lot of IO operations.

We have developed an MPI+OpenMP version, optimizing input data file reading and overlapping these IO operations with computations (signal processing). We run some experiments on a 256-nodes dual-core PC cluster and on a 16-nodes 4-hyperthreaded-cores experimentation PC cluster, both located at SUPELEC. We identified the most efficient configurations, considering the number of input file reading processes, the number of threads per process, the number of processes per node... Finally, we have designed and implemented a deployment tool, in order to hide these configuration issues to signal processing researchers and users.

In a near future, we aim to port this application to ORWL, to experiment, improve and validate ORWL on distributed applications achieving a high ratio of IO compared to computations and classic communications.