This paper describes an implementation of stream sockets for the SHRIMP multicomputer. SHRIMP supports protected, user-level data transfer, allows user-level code to perform its own buffer management, and separates data transfers from control transfers so that data transfers can be done without the interrupting the receiving node's CPU. Our sockets implementation exploits all of these features to provide high performance. End-to-end latency for 8 byte transfers is 11 microseconds, which is considerably lower than all previous implementations of the sockets interface. For large transfers, we obtained a bandwidth of 13.5 MBytes/sec, which is close to the hardware limit when the receiver must perform a copy. Further experiments with the public-domain benchmarks ttcp and netperf confirms the excellent performance of our implementation.

This technical report has been published as

Stream Sockets on SHRIMP. Stefanos N. Damianakis, Cezary Dubnicki and Edward W. Felten, Proc. of 1st Internat. Workshop on Communication and Architectural Support for Network-Based Parallel Computing, February 1997.
(Proceedings available as Lecture Notes in Computer Science #1199.)