Operating System Design Principles for Scalable Shared Memory Multiprocessors

We describe SALSA an operating system that in corporates techniques for achieving scalability in large scale shared memory NUMA multiprocessors We evaluate the e ects of cache organization and caching policy on latency hiding via rapid thread switching With write back set associative caches we demon strate signi cant improvements in program perfor mance with latency hiding when cache miss latency is high even with lowmiss rates SALSA exploits clustering in a NUMA system and provides primitives for hierarchical data placement and thread scheduling We show that proper data placement can double performance in Willow a hierarchical shared memory architecture SALSA also provides user control over memory allocation for ne tuning a pro gram s memory requirements which is shown to im prove program performance by up to