Infiniband for Christmas -- It's a great concept, but will the Infiniband implementation be all it's cracked up to be?
Art Wittmann
Technology has a finite life span-that's just the way it is. The most recent technology marked for retirement is PCI. It likely will continue its tour of duty in desktop systems for some time to come, but its days are numbered in the server room. Just as we've moved from shared to switched backbones on our networks, so are we making the same transition at the server backplane. Venerable, old PCI just isn't cutting it on servers anymore for many reasons, ranging from data-transfer rates to reliability and flexibility. I say we give PCI a place in the techno hall of fame as one of the best standards ever written for computing hardware and leave it at that.
Infiniband, as you probably don't recall, was born after warring factions pushing different bus strategies realized that two standards wouldn't be good for anyone. That was a rare moment of rationality and one for which we should be thankful. It's been almost a year since the Infiniband trade association was formed with charter members from both the Unix RISC camp and the Intel system vendors. Since then most of the industry, including all the major networking hardware vendors, has joined. The specification might be finalized by the time you read this, and-true to form in our impatient industry-chipmakers are turning out parts even though they still don't have a final spec. But that's not as bizarre as it sounds: The final stage of creating a specification primarily involves wrangling about whether the text of the specification correctly communicates already agreed-upon concepts.
Around Labor Day, both Intel and Lucent announced chips critical to implementing the new switched "channel-based" architecture. If you're thinking, "Channel? You mean like microchannel?," the answer is "Good god, no!" But if you're thinking, "Channel! Like mainframe I/O channels!," you're on the right track. The effort focused on bringing the concepts proven in the mainframe world to the rest of the server world. The oft-maligned mainframe has incredible technology in it, and the one realm where mainframes shine very brightly is in I/O processing. Mainframe I/O is so good you'll often find mainframes running the storage farm that feeds supercomputers.
There are a few key concepts in Infiniband that really make it important. As I've written before, separating storage-and even network processors-from the CPU goes a long way toward creating a manageable data center. Infiniband does that by allowing relatively long cable runs, and by pushing server design away from the all-encompassing refrigerator-sized boxes we often see. It also requires the use of channel processors, devices that off-load virtually all bus processing from the CPU. Interrupting the main CPU every time a bus device needs attention has a large impact on overall CPU performance. Most CPUs just weren't designed to handle I/O interrupts all that well and, unfortunately for most of us, Intel CPUs tend to be particularly bad at it.
One of the nicest aspects of Infiniband is that it's designed to support connects of one, four or 12 serialized data streams (running at 2.5 Gbps, which has already been criticized in some circles as too slow). The result is a healthy range of performance and price points without introducing a ton of complexity to the specification. The concepts are definitely good-but good concepts don't always result in good products. Many of us hope that the implementation of these concepts is equally good and that version 1.0 hits the ground running. With such wide-spread support, there's not much doubt you'll be considering Infiniband technology within the next 12 months. If you want to see the first offerings, pack your gambling shoes and head for Spring 2001
NetWorld+Interop.
techweb.com |
