Data Race Detection With Pending Period Representation

With the development of computer technology,parallel programing has become more and more integrated into daily software development as an effective means to increase the efficiency of programing.Although parallel program brings a lot of convenience to the running of programs,it also brings some troubles.Data race is a typical problem that often appears in parallel program but is difficult to be found.Some existing data race detection tools can detect data race precisely,however,they are inefficient and unscalable to be applied in real production.Pending Period based race detection Approach(PPA)detects data races through global clock.This algorithm uses the consistency of the physical clock instead of logic clock to record event orders to reduce high maintenance consistency and increases the program's scalability.In the meanwhile,instead of the way that records the ID of locks to divide different locks,there is no need to distinguish the locks in PPA which further more improves the detection efficiency.In addition,PPA uses local variables to hold the pending periods which can effectively reduce the overhead of frequent access and modify the global variable when accessing the same memory location.Our experimental results on a wide variety of programs show that our approach outperforms state-of-the-art with 53.72%(vs.FastTrack),41.05%(vs.ThreadSanitizer)and 21.01%(vs.IFRit)slowdown improvement on average,which can be more significant as the number of threads is increasing.Compared with overlap-based algorithm IFRit,we also detect more races.At the same time PPA can also use sampling to reduce overhead.Compared with Pacer sampling algorithm,PPA is more efficient(with up to 31.28% improvement in the case of 10% sampling rate).