In existing multicomputers, software overhead dom inates the message-passing latency cost. Our research on the SHRIMP project at Princeton indicates that appropriate network interface support can significantly reduce this software overhead. We have designed two network interfaces for the SHRIMP multicomputer. Both support virtual memory mapped communication allowing user processes to communicate without doing expensive buffer management, and without using sys tem calls to cross the protection boundary separating user processes from the operating system kernel. This paper describes and compares the two network in terfaces, and discusses performance tradeoffs between them.
展开▼
