On the implementation and effectiveness of autoscheduling for shared-memory multiprocessors

Abstract

This thesis addresses the problem of implementing the autoscheduling model of computation on conventional shared-memory multiprocessors. In autoscheduling, the partitioning and scheduling of computations for parallel execution are performed by means of drive code injected by the compiler at the entry and exit points of each schedulable unit (tasks).

A prototype autoscheduling compiler that generates autoscheduling code for real and abstract multiprocessors was implemented. The general organization of this compiler and generated code are discussed in this thesis. The run-time library used by the executable autoscheduling code is also described. Major implementation problems include the execution of the actual scheduling operations, the organization of the task queue, granularity control to adjust the level of parallelism exploited, cactus-stack support, parallel loop implementation, support for data distribution, and execution on a time-variant partition of physical processors.

The correctness and performance of the autoscheduling code generated by the compiler were verified through actual measurements on a real multiprocessor, program level execution-drive simulation, and instruction level simulation. The results demonstrate the feasibility of an autoscheduling compiler and its ability to exploit new levels of parallelism on shared-memory multiprocessors.