This work was carried out under the supervision of

This dissertation addresses three major challenges in designing high-performance concurrent software for modern multicore processors: realizing the potential parallelism of the problem, efficiently dealing with synchronization tasks, and overcoming system-level bottlenecks. Realizing parallelism Obtaining a scalable solution – in which performance improves as close as possible to linearly with the number of threads – can be difficult even for an inherently parallel problem. We demonstrate this using the rendezvous problem, in which threads of two types, consumers and producers, show up and need to be matched each with a unique thread of the other type. The parallel nature of rendezvousing allows each thread to look for a match independently of the other threads. Yet existing algorithms often waste work by attempting to match threads of the same type and by allowing a thread to idly wait until another thread attempts to match with it. We describe an efficient rendezvousing system which avoids these problems, so that every execution cycle goes towards making a successful match, leading to a significant performance improvement. Based on our rendezvousing system, we also construct a highly scalable and competitive concurrent stack implementation (Chapter 4)