Adaptive techniques for improving performance of hardware transactions on multi-socket machines

Abstract

Socket scheduling modes may prevent non-uniform memory access effects from negatively affecting performance of synchronization mechanisms utilizing hardware transactional memory. Each mode may indicate whether a thread may execute a critical section on a particular socket. For example, under transitional lock elision, locks may include a mode indicating whether threads may acquire or elide the lock on a particular socket. Different modes may be used alternately to prevent threads from starving. A thread may only execute a critical section on a particular socket if allowed by the current mode. Otherwise, threads may block until allowed to execute the critical section, such as after the current mode changes. A profiling session may, for a running workload, iterate over all possible modes, measuring statistics pertaining to the execution of critical sections (e.g., the number of lock acquisitions and/or elisions), to determine the best performing modes for the particular workload.