Fine-grain parallel processing on commodity platforms

We present a tight integration of a user-level thread scheduler and a zero-copy messaging system that has been designed and optimized for scalable and efficient fine-grain parallel processing, on commodity platforms, with support for fault-tolerance. The system delivers most of the performance of the underlying communication hardware to a multi-threaded application level, while introducing little CPU overhead. This is demonstrated by a performance analysis of an implementation using off-the-shelf commodity products: PCs, running the Linux operating system, equipped with fast and gigabit Ethernet network interface cards. (21 refs).