|
|||||
Infiniband to get one more tweak before systems ship
Infiniband to get one more tweak before systems ship SAN MATEO, Calif.The Infiniband specification needs one more turn of the crank before the high-speed interconnect technology is ready to ship in systems next year. Version 1.1 of the spec is expected to be complete in September, then servers and other systems that use the gigabit serial I/O are expected to roll out in the middle of 2003. The new version of Infiniband marks a significant reworking of the software infrastructure of version 1.0.a released by the Infiniband Trade Association last June, itself a rework of the 1.0 spec announced at a developers conference in October 2000. The latest version represents not so much a slip in the schedule as a normal process of ironing out high-level software issues resulting from efforts to prototype first-generation Infiniband hardware. Version 1.1 brings support for multicasting, a fresh implementation of subnet manager and subnet administration features as well as a clean up of a number of errata in the previous version. An annex on a sockets direct protocol, just approved by the Infiniband steering committee, will also be part of the new spec. "We've fixed a lot of problems, filled in a lot of holes and refined a lot of ideas that didn't quite pan out when we went to implement them," said one Infiniband developer who asked not to be named. "The industry will be [rolling out products based] on 1.1. It's a good stable release that will last for a number of years. It's what we will base products on." The new version is not expected to have any impact on semiconductor-level Infiniband products. Issues impacting hardware were largely defined by the end of last year, setting up the roll out of chips and boards this year. However, full system-level products are expected to wait for the completion of the new spec and development of software based on it before they ship in the second half of 2003. APIs to pave way for Unix< BR> Separately the Interconnect Software Consortium (ICSC), a de facto standard organization, is drafting the application programming interface specification that will cover high-speed interconnects, including Infiniband. The APIs include a spec supporting asynchronous socket extensions under Unix, a remote direct memory access API covering Infiniband and Gigabit Ethernet, and a fabric-management API. Developers expect the three APIs, which could boost the performance and usability of Infiniband in a Unix environment, will be complete by November. Companies working on those APIs include Compaq, Hewlett-Packard, IBM, Intel, Sun and a handful of Infiniband developers, including Vieo Inc. and Lane 15 Software. A separate group called the Direct Access Transport Collaborative is working in parallel with the ICSC to define kernel and user-level APIs for interconnects such as Infiniband across multi ple operating systems. The APIs will help database companies and other middleware software makers develop products that run on systems using Infiniband or other new interconnect technologies. The software work is seen as a key underpinning to getting full-fledged Infiniband systems deployed to end users. "We expect to see some Infiniband product worthiness this year so that we could take an Infiniband card and plug it into a PCI-X slot in one of our servers, but it will be 2003 before an Infiniband-centric server ships from Dell," said Joe Sekel, a lead server architect at Dell Computer Corp. Proponents see Infiniband as a Holy Grail of high performance interconnects that could become a primary means for data center servers to link to each other, to storage devices and to the network. That could help clean up a mess of separate Ethernet, Fibre Channel and proprietary clustering cables in today's data centers, they say.
|
Home | Feedback | Register | Site Map |
All material on this site Copyright © 2017 Design And Reuse S.A. All rights reserved. |