Burt, that is just one of a hundred profoundly important posts Albert has filled this thread up with on and off for the better part of two years. And he probably still wonders why he got booted from SI.??
Article...Giganet's Cluster LAN will vie with Fibre Channel and P2100 PC interconnect rides on Intel's bus strategy
April 14, 1998
Electronic Engineering Times: Concord, Mass. - Hitching its star to the still unannounced plans of Intel Corp., a small company based here today will roll out an interconnect technology that it hopes will sweep across the high end of the PC I/O world. Giganet Inc. is betting that its Cluster LAN is compatible with an interconnect under development at Intel-one it hopes will eventually spread across the motherboards of high-end servers and workstations as a server, storage and system link.
Giganet's hybrid technology borrows from Fibre Channel, asynchronous transfer mode and the Virtual Interface (VI) specification to define a $795 interconnect board that delivers data rates of 1.25 Gbits/second with latency of less than 8 microseconds. A network switch will follow by midyear.
The startup shares with Intel a belief that a similar interconnect the CPU giant is now defining could ultimately supplant Fibre Channel and a new interface dubbed P2100 for linking networks of servers and storage devices. It could also see use as an internal system bus.
Fibre Channel and P2100 advocates have criticized the relative vaporware of a Gbit serial bus Intel has discussed so far in general terms (see March 9, page 1). But critics and backers alike are quick to concede that Intel has the market prowess to redefine server I/O and networking, if that is indeed its goal.
"What people decide to integrate into the motherboard going forward will determine what the interconnect will be," said David Follett, chairman and chief executive of Giganet. "Intel is in a very dominant position in the market, and they have the ability to drive certain standards."
Giganet has a history of collaboration with Intel in interconnect technology. Several years ago, the company helped develop an OC-12 (622-Mbit/s) ATM interconnect board for Intel's Paragon supercomputer under a U.S. Department of Energy program. Follett said the interconnect board being launched this week is essentially a streamlined version of that programmable-logic-packed board, shrunk to a single ASIC on a half-size PCI card.
Key to the ASIC are algorithms for which Giganet is seeking patents that make the interconnect appear as 1,024 unique interfaces, each of which can be individually addressed by separate applications. "Apps can talk to the hardware directly," rather than through complex protocol stacks on top of an operating system, Follett said. "That leads to tremendous efficiencies."
The interconnect adopts the standard Fibre Channel scheme known as FC0 wiring- the same technique Intel is said to be using in its Gbit serial bus. Giganet layers ATM cells and framing mechanisms on top of that. In addition, it has implemented natively in its silicon the Virtual Interface architecture, a high- level clustering API developed by a broad group of vendors headed by Compaq, Intel and Microsoft.
"The VI architecture is a central piece of this, and it is technically the most complex part of it to develop," Follett said. The VI spec uses virtual addressing so that hardware is responsible for data transfers, shielding software from tracking data addresses that generate significant overhead in a message-passing environment.
The interconnect will initially find use as a link between servers in high- availability clusters using Microsoft's Wolfpack clustering under Windows NT, Follett said. Giganet also expects it to see use as a server-to-disk controller where Fibre Channel is used lower down in the storage-network hierarchy as a disk-to-disk link. "We are trying to pick up the heavy-traffic regions between processors and disks," he said.
Fibre Channel is establishing a beachhead in storage-area networks (SANs). But Follett called Fibre Channel "more of a disk-drive technology. At the high end, it lacks some of the flow-control and queuing mechanisms needed to scale to SANs."
Another contender in this space is IEEE P2100, originally called 1394.2, an interconnect enhanced for clustering that is near balloting as an IEEE standard. Sun Microsystems Inc. is helping drive the standard and is said to be well along in developing ASICs to implement the interface for its own servers.
P2100 is initially being defined as a 1.25- and 2.5-Gbit/s serial bus with target latencies in the range of 2 to 5 microseconds. Clustering enhancements include the use of dual counter-rotating rings for reliability, error-recovery features that preserve low latencies and automatic suppression of dual messages on the bus. It too uses FC0 wiring.
Despite the work on P2100, one engineer who asked not to be named said the VI ideas are rapidly catching on and that Intel could leapfrog both P2100 and Fibre Channel given its market clout. The fact that major database companies are gearing up to support VI will push others to adopt it, rather than create a performance-draining software translation layer for the API. Some P2100 backers are now looking at supporting VI, he said.
An engineer from a mainstream PC company, who also asked not to be named, said the biggest problem is that Intel is keeping details of its bus plans secret- even from top customers. The plan "is not being very openly discussed, " the second engineer said. "To me this is a bad thing for the industry."
Intel and a small group of companies are developing an initial spec for the Gbit link. It may be available in time for the first Merced-based servers, said Justin Rattner, director of Intel's Server Architecture Labs. Once an initial spec is developed, a larger group will review it before it is made public. The bus will borrow ideas from VI, I2O and other efforts, Rattner added.
Perhaps the camp with the most to lose if Intel's architecture takes off is the broad group of chip, board and systems vendors banking on Fibre Channel. The first engineer called Fibre Channel "a disaster," because it involves a complex array of profiles and classes of devices, as well as services such as VI, ATM and IP that working groups are layering on top of it. "There are interoperability problems up the wazoo," the engineer said.
Dal Allan, president of ENDL Inc. (Saratoga, Calif.), a consulting firm that works with the Fibre Channel Loop Community, brushed off the criticism. "I have yet to find any interoperability problems in the real world," he said. "We knew it would take us five years to transfer from a parochial, proprietary world to an open, interoperable standard, and we are just in the first year."
Allan defended the standard as flexible, noting that as many as six profiles may ultimately be defined for the interconnect. And two separate working groups are now taking different approaches to layering support for the VI spec on top of Fibre Channel
"The market is developing in many directions right now, and it's not clear to me Fibre Channel, the monolith, is going to exist," Allan said.
_____________________________________________________________________
Regards, Michael |
