Data-race detection in transactions-everywhere parallel programming

Abstract

This thesis studies how to perform dynamic data-race detection in programs using "transactions everywhere", a new methodology for shared-memory parallel programming. Since the conventional definition of a data race does not make sense in the transactions-everywhere methodology, this thesis develops a new definition based on a weak assumption about the correctness of the target program's parallel-control flow, which is made in the same spirit as the assumption underlying the conventional definition. This thesis proves, via a reduction from the problem of 3cnf-formula satisfiability, that data-race detection in the transactions-everywhere methodology is an NP-complete problem. In view of this result, it presents an algorithm that approximately detects data races. The algorithm never reports false negatives. When a possible data race is detected, the algorithm outputs simple information that allows the programmer to efficiently resolve the root of the problem. The algorithm requires running time that is worst-case quadratic in the size of a graph representing all the scheduling constraints in the target program.