Time adaptation for parallel applications in unbalanced time sharing environment

Time adaptation is very significant for parallel jobs running on a parallel centralized or distributed multiprocessor machine. The turnaround time of an individual job depends on the turnaround time of each of its processes. Dynamic load balancing for unbalanced time sharing environment helps to equally distribute the work load among the available resources, so that all processes of a single job end almost at the same time, thus minimizing the turnaround time and maximizing the resource utilization.; In this thesis we propose and implement an approach that helps parallel applications to use our library so that it can adapt in time dimension (if running in a time sharing environment) without changing the space allocation. This approach provides an interface between application, monitoring information, the job scheduler and a cost model that considers application, system and load-balancing information. This interface allows binding of different adaptation approaches for synchronous adaptation and semi-static remapping. We also determined job types for what this approach is suitable and at the end we present results from our test run on a 16-node cluster with synthetic MPI programs and a time adaptation approach, demonstrating the gain from our approach. In this work, we make extension of existing ATOP [11] work. We directly use their over partitioning strategy. But unlike ATOP, applications can use our adaptation library and adapt dynamically. We also adopted the dynamic directory concept used in SCOJO [8].