Abstracts Engineering

How software prefetching affects transactional memory applications with high commit ratio

by Max Gabrielsson

Transactional Memory is a concurrency control model that allows programmers to write code that access shared data consistently by marking sequential multi-operation regions as atomic transactions. While transactions make programming easier, the lack of progress guarantees as well as the cost of re-executing a conflicting transaction leaves room for improvement. Transactional Memory Decoupled Access Execute (TM-DAE) is an experimental compilation technique that aim to minimize the time spent inside a transaction by inserting preceding access phases which performs software prefetching of the required data in order to decrease the probability of a transaction to fail. TM-DAE has thus far been successful in reducing aborts, but has also been deemed to introduce too great of a performance overhead to be practical. In this project we have developed a technique to apply TM-DAE selectively and investigated the effect TM-DAE has when applied on transactional memory applications selectively. We have hypothesized whether there is a reliable way to identify transactions that will benefit from TMDAE, and defined two metrics to do so, the Commit-Conflict Ratio (CCR) and Commit-Abort Ratio(CAR). Although we have done extensive visualization and discussion of our findings, we conclude that the behavior of an individual transaction is too idiosyncratic for our methods to reliably predict the effect TM-DAE will have. We also discuss two alternative directions for future research to make TM-DAE viable: one based on further static analysis, and another involving a reactive runtime based approach.