Section: New Results

Topology Aware Malleability of MPI programs

Current parallel environments aggregate large numbers of computational resources with a high rate of change in their availability and load conditions. In order to obtain the best performance in this type of infrastructures, parallel applications must be able to adapt to these changing conditions.

In collaboration with Universidade da Coruña, Spain, we have worked on automatically and transparently adapt MPI applications to available resources is proposed. The solution relies on application-level migration approach, incorporating a new scheduling algorithm, based on TreeMatch and Hwloc, to obtain well balanced nodes while preserving performance as much as possible.

The experimental evaluation shows successful and efficient operation, with an overhead of less than 1 second for the proposed scheduling algorithm, and of only a few seconds for the complete reconfiguration, which will be negligible in large applications with a realistic reconfiguration frequency.