SCI SOCKET - A Fast Socket Implementation over SCI

In this paper we introduce a very low latency implementation of the Berkeley Sockets interface on top of the Dolphin SCI interconnect. Average one byte latencies as low as 2.31 microseconds are achieved. To our knowledge, this is the fastest sockets implementation in terms of latency. SCI SOCKET combines outstanding performance and seamless integration with legacy applications. This enables networking applications to transparently exploit the capabilities of SCI without being modified or even recompiled. We describe major design decisions regarding transparent integration with legacy sockets and data transfer protocols. Keywords—Requirements for HPC Systems Software, BSD Sockets, Low latency, SCI, User level communication.