Abstract

This paper describes the scheduling issues speci c to the Tera MTA high performance shared memory multithreaded multiprocessor and presents solutions to classic scheduling problems. The Tera MTA exploits parallelism at all levels, from ne-grained instruction-level parallelism within a single processor to parallel programming across processors, to multiprogramming among several applications simultaneously. Consequently, scheduling of resources occurs at many levels, and managing these resources poses unique and challenging scheduling concerns. This paper outlines the scheduling algorithms of the user level runtime and operating system and describes the issues relevant to each. Many of the issues encountered and solutions proposed are novel, given the multithreaded, multiprogrammed nature of our architecture. In particular, we present an algorithm for swapping a set of tasks to and from memory that achieves minimal overhead, largely independent of the order in which tasks are swapped.

Citations