What Really Makes Transactions Faster?

There has been a flurry of recent work on the design of high performance software and hybrid hardware/software transactional memories (STMs and HyTMs). This paper reexamines the design decisions behind several of these stateof-the-art algorithms, adopting some ideas, rejecting others, all in an attempt to make STMs faster. The results of our evaluation led us to the design of a transactional locking (TL) algorithm which we believe to be the simplest, most flexible, and best performing STM/HyTM to date. It combines seamlessly with hardware transactions and with any system’s memory life-cycle, making it an ideal candidate for multi-language deployment today, long before hardware transactional support becomes commonly available. Most important of all however were the results we derived from a comprehensive comparison of the performance of nonblocking, lock-based, and Hybrid STM algorithms versus fine-grained hand-crafted ones. Contrary to our intuitions, concurrent code generated in a mechanical fashion using our TL algorithm and several other STMs, scaled better than the hand-crafted fine-grained lock-based and lock-free data structures, even though their throughput was lower. We found that it was the lower latency of the hand-crafted data structures that made them faster than STMs, and not better contention management or optimizations based on the programmer’s understanding of the particulars of the structure. This holds great promise for future mechanical generation of concurrent code using hardware transactional support.