SimMatrix: Simulator for MAny-Task computing execution fabRIc at eXascales

Exascale computers will enable the unraveling of significant scientific mysteries. Predictions are that by 2019, supercomputers will reach exascales with millions of nodes and billions of threads of execution. Many-task computing (MTC) is a new viable distributed paradigm for extreme-scale supercomputing. The MTC paradigm can address four of the five major challenges of exascale computing, namely concurrency, resilience, heterogeneity, and I/O and memory; this work specifically addresses the first three major challenges. This paper presents a new light-weight and scalable discrete event simulator, SimMatrix, that enables the exploration of distributed scheduling for MTC workloads at exascale levels with up to 1 million nodes and 1 billion cores. SimMatrix is validated against real MTC workloads executed under Falkon at petascale levels, with 40K nodes and 160K-cores. Centralized scheduling is compared and contrasted to distributed scheduling; this work adopts work stealing, as an efficient and scalable approach to distributed load balancing. It explores a wide range of parameters important to understand work stealing at exascale levels, such as number of tasks to steal, number of neighbors of a node, static or dynamic neighbors, and different workloads. Experiment results show that the centralized scheduling saturates at small number of nodes, while the distributed scheduler configured with optimal parameters could scale up to 1 million nodes and 1 billion cores without any explicit upper bound. SimMatrix is light-weight and scalable, having been tested up to 1 billion cores and 10 billion tasks with modest resources (e.g. 200GB of memory and 256-core hours)