Low Latency Inter Thread Communication over Gigabit Ethernet

The rise in network and processing capacity of commodity platforms has made Networks of Workstations an affordable alternative to dedicated parallel computers. However, standard layered communication protocols (e.g., TCP/IP) on commodity operating systems were not designed for high-performance cluster computing. They introduce bottlenecks due to interrupt overhead, memory copying, protocol processing, and latency from inefficient bus and I/O usage. This book identifies these bottlenecks and presents a user-level communication system over Gigabit Ethernet that bypasses the OS entirely. The library integrates with a user-level thread scheduler to offer multi-threaded applications zero-copy messaging via CSP channels. By replacing interrupts with local polling, it achieves 941 Mbps throughput and significantly reduces CPU load compared to kernel-based communication. This book is ideal for researchers in high-performance cluster computing and distributed systems.