System I/O fabric to transform computing architecture model

1999-09-02

Peace has been declared, it was announced yesterday, between the vendors involved in defining a replacement for the PCI bus. In fact, so much attention was paid to the avoidance of battle, that the potential of the proposed technology was missed by most.

The keyword is “fabric”. System I/O is a fabric-based architecture which enables any device to communicate to any other device. Harware switching is used to enable far higher throughput speeds than previously available, from 2.5Gb to 6Gb per second – this compares with a measly 132Mb per second from PCI. This is all well and good, suggesting that PC devices such as processors, network cards and memory will no longer be limited by bus bandwidth (particularly as the 6Gb top end will probably be extended in the future). Two points have largely been missed, however.

The first is that the system vendors are not the only organisations experimenting with switched fabrics. One notable group is the storage community, for whom the switched fibre architecture is key to the SAN strategies of most companies (including StorageTek, Compaq and HP). What interests the storage vendors is the ability to link switches (by fibre) over long distances: currently, the maximum distances between switches are touted as between 10Km and 50Km. How might the System I/O fabric benefit from long-distance interconnect? For example, a single, conceptual device, running a single operating system could in fact be a pair of mirrored devices, each with its own processor, disk and memory. The processor, disk, memory and graphics card become devices in their own right which, with the inclusion of a switched fabric with a fibre interconnect, could be physically positioned anywhere in a 50km radius but which could be configured dynamically to make best use of the resources at a given time.

The impacts go both ways, so…what impact will the System I/O fabric have on the SAN community? Do we need both types of fabric? How will they interact?

A second point is the inclusion of IP version 6 in the System I/O specification. This effectively removes the need to consider system components as part of the same machine, virtual or otherwise. The Internet is currently IP V4, but IP V6 will be included in most network-ready devices in the future. The potential is clear – that the system bus replacement, System I/O, becomes an integral part of the Internet infrastructure. How this will happen is still a matter for speculation, but the potential this has, of moving us into a device-based world in which bandwidth is a forgotten issue, is clear. Given that System I/O is likely to be one of the most important enabling technologies we have seen, it is only to be hoped that the new consortium can come up with a snappier name.

(First published 2 September 1999)