MPICH-G2 Implementation of an Interactive Artificial Neural Network Training

摘要

Distributed Training of an Artificial Neural Network (ANN) has been implemented using MPICH-G2, and deployed on the testbed of the european CrossGrid project. Load balancing, including adapta-tive techniques, has been used to cope with the heterogeneous setup of computing resources. First results show the feasibility of this approach, and the opportunity for a Quality of Service framework. To give an example, a reduction in the training time from 20 minutes using a single local node downto less than 3 minutes using 10 nodes distributed across Spain, Poland, and Portugal, has been obtained.