Research On Memory Race Recording Method For Multicore Parallel Program

In recent years, multi-core parallel program has been widely used. However,the execution of multi-core parallel program is nondeterministic. The problem will make things difficult for reproducing and locating bugs in the process of multi-core debugging, and take trouble to program and debug parallel program. An effective way to solve this problem is deterministic replay, which is based on recording the memory race information accurately. Existing deterministic replay methods recorded large amounts of memory race information, and bring huge space overhead to the multi-core system. How to reduce log size of memory race is a problem to be solved.To reduce log space overhead, this paper designed a hardware-based multi-core memory race recording method. The method used transitive reduction to reduce the memory race which can be deduced by the recorded memory race and reduced the number of the recorded memory race based on the initial log; and then removed two adjacent memory race in the same direction and replaced by an artificial memory race to further reduce the number of memory race; and recorded the difference between the memory race instead of the memory race to optimize the size of a single memory race. This method reduced the log size by reducing the number of memory race recorded and optimize the size of a single memory race. In addition,this paper also gave a hardware-based description of the algorithm, and implemented the memory race recording method based on MESI directory-based coherence protocol.Finally, this paper built Gem5 multi-core simulation platform, loaded Linux operating system in full system mode, and selected the PARSEC parallel benchmarks set to test the function of memory race recording and the performance on log size. The results showed that the method in this paper could record memory race correctly, and reduced the log size.