İyileştirilmiş paylaşımlı küme kullanımı için melez iş çizelgelemesi

In this thesis, We investigate the models and issues as well as performance benefits of hybrid job scheduling over shared physical clusters. Clustering technologies that are compared include MPI, Hadoop-MapReduce and NoSQL systems. Our proposed scheduling model is above the operating system and cluster-middleware level job schedulers and operating system level schedulers and it is complementary to them. First, we demonstrate that we can schedule MPI, Hadoop and NoSQL cluster-level jobs together in a controlled-fashion over the same physical cluster. Second, we find that it is better to schedule cluster jobs with different job characteristics together (CPU vs. I/O intensive) rather than two or more CPU intensive jobs. Third, we describe the design of a greedy sort-merge scheduler that uses the learning outcome of this principle. Up to 37% savings in total job completion times are demonstrated for I/O and CPU-intensive pairs of jobs, but up to 50% savings (or 2x speedup) is theoretically possible. These savings would also be proportional to the cluster utilization improvements, if there are jobs waiting in the queue. At the end of the thesis, we also discuss potential power-energy savings from hybrid job scheduling.