SCHEDULING AND PAGE MIGRATION FOR MULTIPROCESSOR COMPUTE SERVERS

Several cache-coherent shared-memory multiprocessors have been develop ed that are scalable and offer a very tight coupling between the proce ssing resources. They are therefore quite attractive for use as comput e servers for multiprogramming and parallel application workloads. Pro cess scheduling and memory management, however, remain challenging due to the distributed main memory found on such machines. This paper exa mines the effects of OS scheduling and page migration policies on the performance of such compute servers. Our experiments are done on the S tanford DASH, a distributed-memory cache-coherent multiprocessor. We s how that for our multiprogramming workloads consisting of sequential j obs, the traditional Unix scheduling policy does very poorly. In contr ast, a policy incorporating cluster and cache affinity along with a si mple page-migration algorithm offers up to two-fold performance improv ement. Four our workloads consisting of multiple parallel applications , we compare space-sharing policies that divide the processors among t he applications to time-slicing policies can achieve better processor utilization due to the operating point effect, but time-slicing polici es benefit strongly from user-level data distribution. Our initial exp erience with automatic page migration suggests that policies based onl y on TLB miss information can be quite effective, and useful for addre ssing the data distribution problems of space-sharing schedulers.