Two-level processor scheduling for multiprogrammed NUMA multiprocessors

Non-uniform memory access multiprocessors have potential gains for achieving high performance. These machines are frequently used as computation servers with multiple parallel processes (we use this word instead of jobs) executing at the same time. In such environments, the efficiency of a parallel process can be significantly affected by the processor scheduling strategy. LA two-level scheduling strategy, which is a kind of space-multiplexing scheduling strategy, is a promising one for large-scale multiprocessors. This paper describes the structure and interface of the two-level scheduling strategy. We also propose basic alternatives in this strategy: as alternatives of handling free processors, (1) running-process-first scheme and (2) waiting-process-first scheme; as alternatives of handling idle processors, (1) idle-processor-hold scheme and (2) idle-processor-release scheme. These alternatives are examined through simulation experiments.<>