USENIX 2nd Symposium on
OS Design and Implementation (OSDI '96)
An implementation of the Hamlyn sender-managed interface architecture
Greg Buzzard,
David Jacobson,
Milon Mackey,
Scott Marovich, and
John Wilkes
Hewlett-Packard Laboratories
Abstract
As the latency and bandwidth of multicomputer interconnection fabrics
improve, there is a growing need for an interface between them and
host processors that does not hide these gains behind software
overhead. The Hamlyn interface architecture does this. It uses
sender-based memory management to eliminate receiver buffer overruns,
provides applications with direct hardware access to minimize latency,
supports adaptive routing networks to allow higher throughput, and
offers full protection between applications so that it can be used in
a general-purpose computing environment. To test these claims we built
a prototype Hamlyn interface for a Myrinet network connected to a
standard HP workstation and report here on its design and
performance. Our interface delivers an application-to- application
round trip time of 28us for short messages and a one way time of
17.4us + 32.6ns/byte (30.7MB/s) for longer ones, while requiring fewer
cpu cycles than an aggressive implementation of Active Messages on the
CM-5.
|