Dynamic Variant Rank HEFT Task Scheduling Algorithm Toward Exascle Computing

Abstract

The accomplishment of Exascale systems in the next few years will pose the need for deep hierarchical parallelism and a high degree of heterogeneity, in addition to the expected component failures. Moreover, it will be constrained by energy limitation and power consumption. As a result, Exascale scheduling algorithms need to be robust in balancing the workload consistently, reliably and efficiently amongst the processors. It is also required to achieve the lowest overhead complexity, and, at the same time, decrease the communication costs as low as possible. In this paper, a task scheduler framework is introduced. The framework consists of three parts, the first of which is the resource monitor, the second the task scheduler and the third the dispatcher. The resource monitor explores the resources in the system dynamically, collects the computing resources metadata, and updates the metadata in a continuous manner. The task-scheduler schedules the tasks based on an improved version of the Heterogeneous Earliest Finish Time (HEFT) heuristic, a directed acyclic graph (DAG) scheduling algorithm. The dispatcher receives tasks list from scheduler to allocate tasks on a resources based on HEFT insertion method and constrained. The HEFT algorithm is compatible efficiently for heterogeneous systems and it improves operations without increasing the time complexity. In this paper, our improvement for HEFT algorithm is introduced. The improved algorithm is based on two steps: a static step which uses various schemes to prioritize the nodes (tasks) and a dynamic one. The dynamic step has the ability to allocates tasks that are scheduled statically and is able to receive and allocate new tasks dynamically. The algorithm we named: Dynamic Variant Rank HEFT algorithm (DVR HEFT), is implemented. A Random job generator is implemented to conduct a comparison study to evaluate DVR-HEFT algorithm against some of the state-of-the-art algorithms, specifically HEFT, PEFT, and Lookahead. Our results indicate that DVR-HEFT is able to reduce the scheduler’s makespan in HEFT algorithm without increasing the algorithm’s time complicity.