Infiniband is poised to become the biggest improvement in midrange-to-high-end computing in years, but in spite of the specification's overwhelming vendor support, the arrival of Infini Band doesn't signal the imminent extinction of PCI.
Late last month, the InfiniBand Trade Association announced the release of the InfiniBand Architecture Specification 1.0, which was agreed upon by more than 200 companies, including superheavyweights IBM, Intel, Microsoft and Sun Microsystems. Vendors will harness the new architecture to create highly scalable servers that employ a switched-fabric design, replacing existing bus arrangements, in which messaging traffic and storage traffic run on the same wire.
Prototype products are expected early next year, but eWeek Labs doesn't expect InfiniBand to offer significant real-world returns on investment (even for early adopters) until the second half of next year, when Intel's McKinley processorâ€"which will have hot-plug processor supportâ€"is expected to ship. McKinley will follow Intel's Itanium chips, which are expected to hit the market next March.
Mature InfiniBand products probably won't be available before early 2002.
Scalability will be the primary benefit of InfiniBand. Industry observers believe the technology will first show up in high-performance database clusters because InfiniBand will provide a standard interconnect for cluster nodes.
Although scalability is extremely important, we believe IT managers shouldn't put off their near-term server farm implementations while waiting for InfiniBand-based products. Like all new technologies, InfiniBand will suffer growing pains, which will most likely include interoperability woes and heavy training demands.
Goes both ways
But when infiniband reaches maturityâ€"and we believe it will, eventuallyâ€"its switched-fabric architecture will support both copper and fiber-optic cabling. This will enable IT managers to easily plug in computing modules via a host channel adapter to boost performance, or add storage units through target channel adapters.
InfiniBand components will be identified within the network via an IPv6-based addressing scheme. Sitting between target channel adapters and host channel adapters in the InfiniBand architecture will be InfiniBand switches that intelligently direct transactions and should run far more efficiently than today's contention-based bus architectures.
In the current specification, InfiniBand will come in 500M-bps, 2G-bps and 6G-bps bandwidth options, the lowest of which will not outpace PCI-X (the evolutionary upgrade of PCI, jointly developed by Compaq Computer, IBM and Hewlett-Packard), which runs closer to 1G bps.
PCI will likely stick around for several years, especially in the low-end server market. InfiniBand/PCI bridges will enable IT managers to preserve the long-term investments they have made in legacy hardware and weave this hardware into new InfiniBand clusters.
InfiniBand/Fibre Channel bridges will also be developed to ensure that SAN (storage area network) hardware investments do not go to waste. After spending the last couple of years touting the benefits of SANs, vendors have a lot of incentive not to force their prize clients to repurchase storage hardware for the sake of InfiniBand.
With InfiniBand poised to become the de facto standard for high-end I/O, vendors that want the inside track for selling InfiniBand hardware, or the software to run on this new platform, should consider becoming members of the InfiniBand Trade Association.
For annual dues of US$9,500, IBTA members will have the earliest access to new specifications and can help influence the direction of the architecture by participating in the association's working groups.
Currently, the IBTA has eight working groups in which members can participate: marketing, applications, electromechanical, link, management, software, compliance and interoperability.
