To amortize the cost of MPI collective operations, nonblocking collectives have been proposed so as to allow communications to be overlapped with computation. Unfortunately, collective communications are more CPU-hungry than point-to-point communications and running them in a communication thread on a dedicated CPU core makes them slow. On the other hand, running collective communications on the application cores leads to no overlap. In this article, we propose placement algorithms for progress threads that do not degrade performance when running on cores dedicated to communications to get communication/computation overlap. We first show that even simple collective operations, such as those based on a chain topology, are not straightforward to make progress in background on a dedicated core. Then, we propose an algorithm for tree-based collective operations that splits the tree between communication cores and application cores. To get the best of both worlds, the algorithm runs the short but heavy part of the tree on application cores, and the long but narrow part of the tree on one or several communication cores, so as to get a trade-off between overlap and absolute performance. We provide a model to study and predict its behavior and to tune its parameters. We implemented both algorithms in the multiprocessor computing framework, which is a thread-based MPI implementation. We have run benchmarks on manycore processors such as the KNL and Skylake and get good results for both performance and overlap.

Non-blocking collectives have been proposed so as to allow communications to be overlapped with computation in order to amortize the cost of MPI collective operations. To obtain a good overlap ratio, communications and computation have to run in parallel. To achieve this, different hardware and software techniques exists. Dedicated some cores to run progress threads is one of them. However, some CPUs provide Simultaneous Multi-Threading, which is the ability for a core to have multiple hardware threads running simultaneously, sharing the same arithmetic units. Our idea is to use them to run progress threads to avoid dedicated cores allocation. We have run benchmarks on Haswell processors, using its Hyper-Threading capability, and get good results for both performance and overlap only when inter-node communications are used by MPI processes. However, we also show that enabling Simultaneous Multi-Threading for intra-communications leads to bad performances due to cache effects.

No abstract