next up previous
Next: High Performance Parallel Up: Available Networking Technologies Previous: Available Networking Technologies

Scalable Coherent Interface (SCI)

SCI technology is based on the ANSI/IEEE SCI standard 1596-1992 and is designed to provide low latency and high bandwidth interfaces for workstation clustering. In particular, this technology can provide latency and, consequently, can efficiently support both shared memory and message passing applications between a variety of workstation types. Bandwidth on individual links is one gigabit-per-second, and the standard can support either ring or switched communications topologies. The SCI standard can be used to create a virtual bus, so that workstation internal bus protocols can be extended between (heterogeneous) clusters of computers.

I found only one company offering SCI products at SuperComputing '94 - Dolphin Interconnect Solutions, a company with main offices located in Norway. Dolphin offers SCI interface cards to interconnect SBus-based or VME-bus-based workstations. In additional, this company offers bridging solutions between:

The SBus product offered by Dolphin has software that provides the options either to use low latency UNIX device drivers accessed directly by applications, or to use the internet protocol (IP) socket-based protocol stack for existing applications. This relationship is described in greater detail above, and is illustrated in figure 4. Due to the fact that this technology can be used for either low-latency bus interconnections or transparent transmission control protocol (TCP)/IP interconnections, figure 3 shows this networking technology as being able to provide LAN connectivity as well as bus connectivity. Dolphin offers a mailbox on the card that can be used for low latency message exchanges, which requires application program modifications to directly send data from memory to the interface card via direct memory access (DMA).

No pricing information was available for this networking technology, however, interconnections are made with relatively low cost shielded twisted pair cables. At these bandwidths, this physical interconnection media surely limits the distances between clustered workstations. Dolphin is examining using ATM-based communications to extend the distances between clustered workstations. If computers are clustered in a ring topology, latency could suffer as a function of the distance (number of network hops) from transmitter to receiver, which is a function of their relative locations on the ring. The SCI protocol can be used in a switched architecture, however, no information on switch availability or pricing was available. In general, this is a limited use networking technology, only to be used in high-performance workstation clusters.



next up previous
Next: High Performance Parallel Up: Available Networking Technologies Previous: Available Networking Technologies



David P. Koester
Sun Oct 22 13:05:27 EDT 1995