Architecture for high performance Knowledge Grid Scheduler

In order to improve the performance of data mining applications, an effective method is task parallelization. The scheduler on Grid plays an important role to management subtasks so as to achieve high performance. We introduce an additional component that we call serializer, whose purpose is to decompose the tasks in the directed acyclic graph (DAG) into a series of independent tasks, and send them to the scheduler queue as soon as they become executable with respect to the DAG dependencies. The experimental result demonstrates that the architecture has good performance.