A Framework for Multiple Parallel Task Graphs (PTG) Scheduler

Abstract

Many applications in scientific computations exhibit both data and task parallelism. Several studies have proved that designing parallel applications using both task and data parallelism is an effective approach than pure data or pure task parallel models. This mixed parallelism achieves both higher scalability and performance. Mixed parallel applications are represented as Parallel Task Graph (PTG), a graph of data parallel tasks. Scheduling such a mixed-parallel application is NP-complete even on a single homogeneous cluster. To maximize resource utilizations and to increase cluster throughput, multiple applications are scheduled concurrently on a cluster. Scheduling multiple applications is challenging as different applications compete for the shared resources and also fairness must be ensured. A new method to perform concurrent schedule of multiple PTGs on a cluster is proposed in this work. Further a complete framework to schedule PTGs submitted at different instants of time and to vary processor allotment for each application during their depending on processor availability is proposed. From simulation experiments, it is observed that the proposed method to schedule multiple PTGs performs better than other methods found in the literature. The suggested scheduler framework to handle online submission of PTGs is proved to be a promising one.