Dodging Non-uniform I/O Access in Hierarchical Collective Operations for Multicore Clusters

The increasing number of cores led to scalability issues in modern servers that were addressed by using non-uniform memory interconnects such as Hyper Transport and QPI. These technologies reintroduced Non-Uniform Memory Access (NUMA) architectures. They are also responsible for Non-Uniform Input/Output Access (NUIOA), as I/O devices may be directly connected to a single processor, thus getting faster access to some cores and memory banks than to the others. In this paper, we propose to adapt MPI collective operations to NUIOA constraints. These operations are now often based on the combination of multiple strategies depending on the underlying cluster topology, with local leader processes being used as intermediate. Our strategy focuses on electing these leaders according to the locality of processes and network interfaces so as to give them privileged network access. We validate our approach on a hierarchical Broadcast operation which brings up to 25% throughput improvement between 64 processes.