Dynamic work-unit slicing for time-constrained job execution in P2P environments

Efficient utilization of resources is one of the key challenges that project administrators face in the dynamic peer-to-peer (P2P) computing environment. The volatile nature of participation by the general public for running lengthy and complex scientific applications is not bounded by any legal obligations, thus, requires following a greedy approach for resource utilization. Generally, the availability statistics of a computational resource and selected set of scheduling algorithms or heuristics are combined together for allocating tasks to the available resources. Berkeley Open Infrastructure for Network Computing (BOINC) is a widely used middleware platform in P2P systems, which uses four interrelated scheduling policies for utilization of available pool of computational resources. The policies used by BOINC mainly depend on the job execution estimates based on the composition of the job and performance results of the considered computing platform obtained using benchmarks. BOINC deploys two traditional synthetic benchmarks: Dhrystone and Whetstone; for measuring the integer and floating-point performance of a considered platform. However, the performance results obtained using these benchmarks show significant variations in results for similar microprocessor, operating system and hardware configuration. These inconsistent results, when used for scheduling, significantly affect the resource-scheduling estimates particularly for time-constrained jobs. This study proposes a novel scheduling policy based on a more consistent and P2P representative benchmark - MalikStone. The policy considers the total availability time of a computational resource, estimated execution time of a work-unit and the available unused time of a computational resource for dynamically slicing large work-unit into smaller work-unit depending on the available unused time of a computational resource. The results have revealed that the policy improved the utilization of available computational - esources by around 10% under the considered experimental settings.