Fair Adaptive Parallelism for Concurrent Transactional Memory Applications
Amin Mohtasham, João Barreto · 2015
Modern parallel machines are likely to run multiple parallel processes together. However, collocating parallel processes in a single machine can easily result in cross-process and cross-thread interferences that can dramatically degrade the system's performance. Such interferences can be mitigated by dynamically adjusting each process' parallelism towards a fair and efficient configuration. We propose a decentralized method for adaptive parallelism for collocated transactional multi-threaded processes. Inspired by well-known results from flow/congestion control mechanisms in communication networks, our technique adopts a hill-climbing strategy that was previously unexplored in the context of adaptive parallelism.