Congestion Management for Hpc Interconnects Using Distributed Routing: Challenges and Solutions

In the era of Exascale computing, the interconnection network is a key component in the High-Performance Computing (HPC) Systems which are expected to interconnect million of processing nodes by the year 2018. It is a prominent challenge for current network technologies such as InfiniBand to achieve high network performance in order not to be the bottleneck of HPC systems. In that sense, Congestion menaces the network performance and, consequently, the overall HPC system operation. This book covers important gaps in congestion management for HPC interconnection networks using distributed routing. It analyzes the drawbacks of proposed solutions, where none of the existing techniques for this type of networks achieves a complete efficiency and scalability. In that sense, the new proposals adapt the RECN basic principles to distributed-routing networks, by means of efficient and scalable techniques which can be divided in two groups: simple techniques which achieve the optimal network performance with the available resources, and completely-efficient techniques which achieve almost the maximum network performance at an affordable cost in terms of network resources.