|M.Sc Student||Ben-Shoshan Yaron|
|Subject||Labeling Fork-Join Task Graphs for Data Race Detection|
|Department||Department of Electrical Engineering||Supervisor||Professor Yoram Moses|
Labeling schemes are a central tool in dynamic data race detection. The space complexity of the labels and the time complexity of comparing two labels are a crucial factor in the time and space complexity of data race detection methods that use that scheme. This work establishes two tight lower bounds on the space complexity of a single label produced by a labeling method in a fork-join run. Consider the following parameters of a fork-join structured run :
- The maximal nesting level of fork-join pairs in r.
- The size of a maximal set containing pairwise parallel blocks.
- The maximal number of fork-join pairs within the same nesting level.
- The maximal number of threads forked form a “fork” operation.
One lower bound, which is given in term of the parameters N, S and L is , and the other, which is given in term of T and L alone, is . We also prove that methods exist whose space complexity meets the lower bounds. The first method to be mentioned under this context is called DTN (Decomposition Tree Numbering), and it matches the lower bound of . Instead of presenting a labeling method that meets the lower bound of , we present a general translation between labeling methods. Assume we are given a labeling method with a space complexity of . The result of translating this method is a new labeling method with a space complexity of . In this manner we conclude that the conversion of DTN meets the lower bound. This work extends, generalizes, and unifies the analysis of labeling schemes presented in a number of previous works.