Scalability Analysis and Evaluation of Divisible Load Scheduling

In this work, we address the problem of the scalability of divisible load scheduling of data parallel workloads (also called arbitrarily divisible workloads) on high performance parallel and distributed computing systems. Divisible load theory offers a linear, deterministic, and tractable model for scheduling arbitrarily divisible workloads, often encountered in scientific applications. With the continuous increasing problem and the system sizes, it is imperative that application scheduling algorithms scale well to leverage the processing capabilities of the high performance computing systems. We conduct an analytical evaluation as well as a simulation-based study of the scalability of divisible load scheduling algorithms, called DLT algorithms, when applied for scheduling two NAS parallel benchmarks, namely, the embarrassingly parallel (EP) and the integer sort (IS) benchmarks onto a target system modeled as a 3-d torus topology. The EP benchmark is computationally intensive and the IS benchmark is communication intensive. Two questions related to the scalability study are addressed, namely, the fastest time to solve the problem and the condition for cost-optimality. A simulation-based study was conducted to support and verify the solutions obtained via the analytical model.