Dynamic routing balancing on InfiniBand network

InfiniBand (IBA) technology was developed to address the performance issues associated with messages movement among Endnodes and computer I/O devices. However, InfiniBand is also widely deployed within high performance computing (HPC) clusters due to the high bandwidth and low message latency attributes it offers to inter-processor communication systems.
An interconnection-network efficient design is mandatory because its great impact on the parallel computer performance. Therefore, a high speed routing scheme that minimizes congestion and avoids hot-spot areas should be included in the network components. We have developed Dynamic Routing Balancing (DRB), an adaptive routing mechanism that balances the communication traffic over the interconnection network. It is based on limited and load-controlled multipath expansion in order to maintain low and bounded network latency. In this work, we propose using DRB as the congestion control mechanism for InfiniBand networks. Experimentation shows that our method achieves significant performance improvement over the original InfiniBand technique which is based on message throttling. An improvement up to 66% for latency and 35% for throughput is achieved for the networks under analysis. Finally, the proposed mechanism use the management model defined in InfiniBand specs, thus full compatibility is provided.
Facultad de Informática