The Power of Scheduling-Aware Synchronization

We present Hydra, a blocking synchronization protocol which significantly outperforms all previous state-of-the-art synchronization algorithms. We experimentally show that the throughput of Hydra is higher than that of CC-Synch, a state-of-the-art synchronization protocol presented in PPoPP ’12, by a multiplicative factor of up to 7.9; this throughput is surprisingly close to the ideal. Hydra implements the combining technique. It also employs user-level threads and schedules them appropriately to ensure their synchronization, as well as to improve performance. We show that by appropriately using user-level threads, we can also get a simple variant of PSim, called PSimX, with highly upgraded performance than that of PSim; PSim is a state-of-the-art wait-free universal construction, first appeared in SPAA ’11. The performance of PSimX, albeit lower than that of Hydra, is also very close to the ideal. Experiments show that by employing user-level threads in other state-of-the-art synchronization protocols, we do not get so big performance advantages as for Hydra