学术文献库

Testing for causality between events in distributed executions is a fundamental problem. Vector clocks solve this problem but do not scale well. The probabilistic Bloom clock can determine causality between events with lower space, time, and message-space overhead than vector clock; however, predictions suffer from false positives. We give the protocol for the Bloom clock based on Counting Bloom filters and study its properties including the probabilities of a positive outcome and a false positive. We show the results of extensive experiments to determine how these above probabilities vary as a function of the Bloom timestamps of the two events being tested, and to determine the accuracy, precision, and false positive rate of a slice of the execution containing events in the temporal proximity of each other. Based on these experiments, we make recommendations for the setting of the Bloom clock parameters. We postulate the causality spread hypothesis from the application's perspective to indicate whether Bloom clocks will be suitable for correct predictions with high confidence. The Bloom clock design can serve as a viable space-, time-, and message-space-efficient alternative to vector clocks if false positives can be tolerated by an application.