Extracting Task-Level Parallelism

Abstract

Automatlc detection of tusk (also referred to as functional. DA(+ unstructured or thread parallelism) at various levels of program granularltv E becornlng mcreasmgly Important for parallehzmg and back-end compders Parallelmlng compilers detect lteration-level or coarser granularity parallehsm which is suitable for parallel computers, detection of parallehsm at the statement or operation-level is essential for most modern microprocessors mcludmg superscalar and architectures In this article ve study the problem of detecting, expressing and cptlnnzmg task-level parallehsm where “task” refers to a program statement of arbitrary granularity Optlmlzmg the amount of functional parallelism (by allowing synchronization between arbitrary nodes) m sequential programs reques the notion of HI term of paths m graphs which Incorporate control and data dependence Precedences have been defined before In a dfferent context: however the defimtmn was dependent on the Ideas of parallel execution and time Ve show that the problem of determmmg precedences statically IS NP-complete Determmmg precedence relationships E useful m finding the essential data dependence show that there a unique mmlmum set of essential data dependence, finding tlms mmlmum set M NP-hard and NP-easy We also propose a heurlst]c algorithm for flndlng the set of esentlal data dependence Static analysls of a program in the Perfect Benchmarks vas done and we present some experimental results. © 1995, ACM. All rights reserved.