The future of system I/0 in the datacenter.

To the casual observer, the system I/O and server interconnect market looks to be in a huge state of chaos. Not so! True, there are more competing technologies hovering around making noise, but the unifying vision of how datacenter infrastructure is evolving has not changed at all. It is Strategic Research Corporation's view that where we are going is still clear and it is only the timeline that is fuzzy. This vision has two elements, the vision of the future of datacenter architecture and the evolution of system I/O.

Step back from the noise and ask the question, "What does the datacenter infrastructure have to become to host the requirements of outward-facing "B-B Internetworking" datacenters and ultimately the "Information Utility Network?" Strategic Research believes that the datacenter of the future looks like Figure 1. Servers and storage have disaggregated and are now interconnected via multiple interoperating fabrics. Central to the new architecture is the SIF, the System Interconnect Fabric. Servers are compute nodes based on blades in an intelligent, high performance, highly scalable, continuous operations fabric. The issues of which protocol have gone away as the SIF carries whichever type of traffic the application demands with intelligent edge devices performing offload functions. Storage is also connected to the SIF via intelligent appliances (more blade servers) who are protocol and I/O type (file or block) agnostic. Disk arrays become 'dumb' again as the intelligence disaggregates into the appliances, causing further commoditization of disk. QoS, Quality of Service, is the driver now, not availability, as in a shared 'continuous operations' environment it is the issues of access and provisioning based on guaranteed QoS-levels that matter the most. What is critical from a systems I/O perspective in this vision is that I/O has to externalize to enable server disaggregation. The path is provided through InfiniBand, IB, technology.

[FIGURE 1 OMITTED]

Now, the second element of confusion currently in the market is the subject of server-class internal I/O (bus and backplane) versus external I/O (SIF). An important shift occurred mid-last-year when IB no longer became considered for a role in "internal I/O" and HyperTransport and Rapid I/O leapt into the gap. Figure 2 illustrates the system I/O evolutionary path we are now on. The key transition is the movement to a serial architecture for the bus, backplane, and interconnect fabric to create a performance platform that scales with the rest of the network infrastructure,

[FIGURE 2 OMITTED]

Going forward, IB owns the external I/O role (the SIF) and PCI the internal. Yes, PCI. PCI owns 95+% of the server bus and backplane market today and it won't go away anytime soon. But, it's a parallel bus technology not serial. This means that while PCI (as in PCI-X, -2X, and maybe--4X) will dominate the parallel bus in PCs and servers for years to come, HyperTransport (HT), Rapid I/O (RIO), and now 3GIO present a new set of internal bus and backplane options as they are released. From a server, datacenter, and application infrastructure perspective, the same "old-rules" still apply. "The migration path which is least painful and maintains the most commonality will win." This is exactly what 3GIO is counting on as the PCI-SIG owns the standard and is creating a smooth migration path from PCI to 3GIO in the 2004 timeframe. From the perspective of the SIF, it doesn't matter much as all three internal I/O buses are designed for IB connectivity.

Put all this in perspective now. First, the emergence of the System Interconnect Fabric based on InfiniBand is a crucial next step in datacenter infrastructure evolution. IB is progressing very well. (Keep in mind that we are only one year into productization). Second, don't count PCI down-and-out or discard 3GIO because it is so far away (2004). The server companies are conservative and will make change slowly and only with great consideration. On the other hand, HT and RIO are already being integrated for important bus and backplane roles in non-server products.

www.sresearch.com

Michael Peterson is President of Strategic Research Corp. (Santa Barbara, CA) and the author of two new books analyzing the evolution of the datacenter and the changes brought about to its infrastructure, "The Future of the Business Network" and "The Future of System I/O".