Accordion Clocks: Logical Clocks for Data Race Detection

Events of a parallel program are no longer strictly ordered as in sequential programs but are partially ordered. Vector clocks can be used to model this partial order but have the major drawback that their size is proportional to the total number of threads running in the program. In this article, we present a new technique called ‘accordion clocks’ which replaces vector clocks for the specific application of data race detection. Accordion clocks have the ability to reflect only the partial order that is relevant to data race detection. We have implemented accordion clocks in a Java virtual machine and show through a set of benchmarks that their memory requirements are substantially lower than for vector clocks.