Trade Press - Spec aims to speed server I/O
By APRIL JACOBS
Network World, 10/23/00
nwfusion.com
The InfiniBand Trade Association this week will deliver
the first specification outlining its high-speed bus
architecture, which is designed to pave the way for
faster communications between clustered servers and
help remove I/O bottlenecks between servers and
network resources.
For customers, InfiniBand will mean no longer having to
plug multiple adapters into servers for network, storage,
and cluster connects. Instead, a connection to an
InfiniBand-compliant switch will provide a virtual fat pipe,
saving users from installing and maintaining complex I/O
configurations. The result is a simpler way of letting
server processors connect to peripherals, storage and
other servers to gain greater performance.
This week's conference will also feature new InfiniBand
products and the formation of a working group that will
be charged with ensuring interoperability between
InfiniBand products.
On the product side, IBM will announce a chip that
powers InfiniBand switches due out late next year.
Texas firm Banderacom plans to introduce a chip
architecture to support InfiniBand storage, server and
network products.
Driving the need for InfiniBand is the 1G byte/sec
bus-speed limitation that even advanced PCI-X-based
servers will have. InfiniBand technology will set up a
switched fabric backplane that can handle 500M
byte/sec to 6G byte/sec links and throughput of up to
2.5G byte/sec.
Using products that support the technology, the number
of addressable devices on a network theoretically could
hit 64,000.
By connecting multiple InfiniBand networks, the number
of devices with addresses is virtually limitless, says Jim
Pappas, a marketing director for the Intel Enterprise
Products Group.
InfiniBand will support copper and fiber-optic cable links
and let entry-level to high-end servers more easily adapt
to transaction-based traffic on the network.
But InfiniBand technology isn't just for servers. A central
InfiniBand I/O switch, which sits on the network to
connect devices, could also be used to cluster servers,
acting as the high-speed interconnect be-tween them.
IBM plans to announce three types of InifiniBand
technology, including an App-lication Spe-cific
Integrated Circuit chip that will power InfiniBand
switches, a host channel ada-pter that connects the
InfiniBand switch fabric network to a server, and a Target
Channel Adpapter designed to connect I/O devices,
such as a network or storage device to the switch fabric.
Banderacom will announce IBandit, an InfiniBand
semiconductor architecture that will provide the
foundation for a family of InfiniBand Channel Adapters.
The comp-any also plans to announce a partnership
with Wind River Systems, which makes software and
services for Internet devices, to develop and supply
InfiniBand transport software that will include support for
Banderacom's IBandit architecture.
Several start-ups, including Cicada Semiconductor and
INH Labs in Austin, Texas, are also developing
InfiniBand technology.
New group formed
Meanwhile, Tom Bradicich, director of architecture and
technology for Netfinity and xSeries servers at IBM, as
well as co-chair of the InfiniBand Trade Association,
says a newly formed workgroup to be introduced at this
week's conference will determine how third-party
products will work together using InfiniBand.
For customers, that means they could cluster servers
from several vendors and deploy multivendor storage and
network devices.
A bright future
Also at the meeting, Vernon Turner, a vice president at
Framingham, Mass., research firm IDC, plans to present
a market forecast for InfiniBand-enabled devices.
He says the total market will reach about $2 billion by
2004. That market represents the hardware opportunity
for suppliers to develop and install InfiniBand
technology, which includes servers, switches and other
devices.
This means that of the approximately six million servers
shipped in 2004, four million could be
InfiniBand-enabled.
But InfiniBand has to co-exist with and help existing
standards that power the industry's network, server and
storage devices.
"InfiniBand's success relies heavily on cohabitating with
the existing fabrics of PCI, Ethernet and Fibre Channel,"
Turner says.
InfiniBand could also boost reliability. For example, if a
server's PCI bus fails, it goes down. But not so with
InfiniBand. Because of the way it is implemented, it can
provide multiple connection paths. If one thread or
connection is lost, others can be accessed, similar to
the way mainframes have access to multiple partitions.
Intel's Pappas says InfiniBand will give users a type of
grid computing, similar to the electrical power grids that
light up the country. If one source becomes unavailable,
another can be accessed to take its place and fill
bandwidth needs.
What InfiniBand will eventually replace is a legacy of
shared bus technology that requires critical server and
network components to share resources, which is not
sufficient for today's data centers.
"As you add more connections, you don't share
bandwidth, you add more. With today's bus
technologies, each card has to take its turn, where
InfiniBand allows each connection to talk at the same
time," Pappas says.
InfiniBand Architecture: Bridge Over Troubled Waters (PDF)
infinibandta.org
InfiniBand white papers
intel.com |
