Benchmarking and Analysis of Variations of Work Stealing Scheduler on Clustered System

Classical work stealing is an efficient dynamic load-balancing technique in shared memory multiprocessor or multicore system. But the performance of the same classical work scheduling on cluster chip multicore is not appreciable. So modification to this is necessary to improve performance. In this paper, we have discussed many earlier proposed modifications, and also proposed some simplistic modifications to suite targeted clustered environment. We have described a methodology to evaluate all the variations of work stealing analytically and experimentally on multiprocessor simulator and on real platform. Our methodology of evaluation include designing of novel parametric synthetic benchmark, which can be used to mimic behavior (or profile) of many real life benchmarks. The designed synthetic benchmark caters a wide range of application profiles to evaluate the design space of both variations of work stealing algorithms and clustered chip multiprocessor. In this work, we found that if the number of available parallelism of the targeted application is higher and data sharing between tasks is high then one of the proposed modification of work stealing (probabilistic based victim search and threshold on size of migratable task) outperform the rest of the modifications.