Hadoop集群公平调度算法的改进

For some specific system and applications, jobs are almost same size, but have clearly defined priorities. The original fair scheduler can't satisfy this need. We propose a improved fair scheduler to control slots preemption based on the job priority which it is only used as job weight in original scheduler. The experiments' results show that the improved scheduler can guarantee that the preemption won't happen between running high-level job and new low-level job and it's always the most important job first completed in a certain time.