Experiences in initiating concurrency software research efforts

Multi-core CPUs are now common in modern computers. To get access to effectively an unlimited supply of compute resources, software programs that have been highly optimized to use a single CPU need to be converted where possible to use concurrency. We have initiated our concurrency software research for performance enhancement on a large-scale system with high throughput and low latency transactions. In this paper, we report our experience, experiments, and results in various aspects of concurrency design and programming, including multi-threaded prototypes, static and dynamic concurrency analysis, future techniques and trends, concurrency experiments, and concurrency design patterns. Based on the concurrency experiments, we achieved at least 80 percent overall performance increases as measured by transaction throughput. As a result, capital expenditures for large scale deployments can be significantly reduced.