Efficient Multi-Processor Scheduling in Increasingly Realistic Models (Brief Summary)

We study the problem of efficiently scheduling a computational DAG on multiple processors. While previous works have mostly studied this problem in rather simple models, we instead consider the well-established BSP model, and extend it with non-uniform memory access (NUMA) effects. This results in a notably more realistic model that captures communication costs, synchronization costs, and the hierarchical structure of modern computing architectures. We develop a range of algorithms to minimize the scheduling cost in this more complex setting: several initialization heuristics, a hill-climbing local search, and different approaches to solve the problem as an Integer Linear Program (ILP). We analyze our scheduling algorithms on a diverse set of real-world computational DAGs to show that they outperforms both academic and practical baselines: without NUMA effects, our scheduler finds solutions of 44%-51% smaller cost on average, and with NUMA effects, it achieves up to a factor 3× improvement. Finally, we develop a multilevel scheduling algorithm for the special case when the problem is dominated by very high communication costs.