Performance Cost Model for Non-DOALL Loops

The thread-level parallelism (TLP) present in DOALL loops is well understood and easy to exploit with modern parallel programming frameworks and tools. Non-DOALL (or DOACROSS) loops also contain some level of TLP, but inter-iteration dependences make it hard to exploit without costly synchronization overhead. In this work, a performance cost model is proposed that uses a DOACROSS loop's characteristic performance parameters (a minimal set of values that can be gathered statically or dynamically) to determine if the loop is worth parallelizing. The performance model is based on both a thorough analysis of the effects of these parameters and supportive experimental results.

Major Advisor: Ben Lee
Committee: Bella Bose
Committee: Thinh Nguyen
Committee: Ted Brekken
GCR: William Warnes