Using Dynamic Broadcasts to Improve Task-Based Runtime Performances

Task-based runtimes have emerged in the HPC world to take benefit from the computation power of heterogeneous supercomputers and to achieve scalability. One of the main bottlenecks for scalability is the communication layer. Some task-based algorithms need to send the same data to multiple nodes. To optimize this communication pattern, libraries propose dedicated routines, such as MPI_Bcast. However, MPI_Bcast requirements do not fit well with the constraints of task-based runtime systems: it must be performed simultaneously by all involved nodes, and these must know each other, which is not possible when each node runs a task scheduler not synchronized with others. In this paper, we propose a new approach, called dynamic broadcasts to overcome these constraints. The broadcast communication pattern required by the task-based algorithm is detected automatically, then the broadcasting algorithm relies on active messages and source routing, so that participating nodes do not need to know each other and do not need to synchronize. Receiver receives data the same way as it receives point-to-point communication, without having to know it arrives through a broadcast. We have implemented the algorithm in the STARPU runtime system using the NEWMADELEINE communication library. We performed benchmarks using the CHOLESKY factorization that is known to use broadcasts and observed up to 30% improvement of its total execution time.