DEC readies High-end OS
Digital sees new Galaxies in high-end computing
By W. David Gardner, EE Times
Maynard, Mass. -- The pioneer of computer clustering, Digital Equipment
Corp., claims it is in the early stages of systems-software work that could
lead to a new way of building high-end computer systems. While few details
about the project--dubbed Galaxies Software Architecture--are available,
one engineer involved in it suggested the work could open up a new debate
in computer architectures.
The heart of the effort appears to be a new capability for launching multiple
instances of an operating system that work cooperatively as if part of a
single computer, an approach that Digital's engineers call Adaptive Partitioned
Multi-Processing (APMP). The technique could create a bridge between two
classes of high-end systems: those that use a shared-memory approach to create
a single system image across multiple processors, and those that take a
shared-nothing approach to building computers that link a number of autonomous
processing nodes via message-passing software.
Digital is not releasing details about Galaxies, in part because the company
says the project is still 12 to 18 months away from launching a product.
However, one Digital engineer agreed to provide a sketch of the effort, which
he said involves "brain surgery on the operating system." The result, he
said, was "the next thing beyond clustering."
"We've got something new," said Stephen Zalewski, a software engineer in
Digital's OpenVMS systems-software group, which is designing Galaxies. "Galaxies
is not just an incremental upgrade to clustering. It's the next computing
model," comparing Digital's APMP concept to other computer architectures
including symmetric multiprocessing (SMP) and NUMA (non-uniform memory access).
Part of the motivation behind Galaxies, Zalewski said, was to find a bridge
between today's architecturally diverse systems, some of which share memory
among processors and others that pass data between distributed systems. "We
wanted to find a way for the end user to run all these models with one solution,"
says Zalewski. "Galaxies is our way of responding."
The first incarnation of the Galaxies architecture will be under the OpenVMS
operating system on Digital's Alpha-based systems. But Zalewski said the
underlying concepts are network- and hardware-neutral and could be applied
to other vendors' systems. "Initially Galaxies will be for OpenVMS, but it's
not necessarily just wedded to OpenVMS," Zalewski said.
Terry C. Shannon, editor of the newsletter Shannon Knows DEC, said there
was a buzz about Galaxies at Digital's recent annual user convention in
Cincinnati, where the company disclosed some of its work to a small group
of customers under a non-disclosure agreement. However, despite the backroom
briefings, even some major Digital customers have only a sketchy picture
of Galaxies, and conflicting information is beginning to spread by word of
mouth and on Internet newsgroups.
Only a few details of the Galaxies software have emerged so far. For instance,
it will at least support basic features like load balancing across distributed
systems. An internal Digital document says that "OpenVMS APMP can load-balance
the resources among the partitions in real time."
Processor, memory or I/O cards could be plugged into or pulled out of a
Galaxies-based computer without rebooting. Reportedly, all partitions in
a Galaxies system will be able to share memory with a command database as
well as cache memory.
"This evolutionary way of using multiple CPUs enables users to partition
a single stem into multiple SMP nodes," said Shannon.
Shannon estimated that Galaxies will support a single system of up to 256
CPUs, 1 terabyte of physical memory and up to 500 I/O ports, to deliver 1
million transactions per second. Sources at Digital say there is no real
limit to the number of nodes that can operate under Galaxies.
The most obvious applications would be supporting large databases for data
mining, data warehousing or transaction processing. Zalewski suggested that
the Galaxies architecture might some day even have a role in low-end systems.
Word of Digital's efforts to propel itself into the forefront of high-end
business computing come none too soon. Although Digital led the industry
in rolling out ways of interconnecting individual systems into high-end clusters,
competitors have been encroaching.
In a recent review of cluster computing, research firm Technology Business
Research Inc. (Hampton, N.H.) examined various features including scalability,
availability, configuration flexibility, connectivity and applications support.
The survey concluded that Digital's OpenVMS clusters still lead over products
from Hewlett-Packard, IBM, NCR and Sun Microsystems.
"If you look at history, Digital has had the clustering expertise," said
Jim Garden, director of technical services at TBR, "but either the company
didn't market properly or the market wasn't ready. Now the market is ready
for cluster computing, but Digital has slipped from the No. 2 spot in computing
to the No. 9 spot."
Microsoft Corp.'s cluster technology for Windows NT, called Wolfpack, is
expected to ship in an early-release version later this year. However, the
software is still in its formative stages and currently only addresses
availability issues across a small number of systems.
As NT grows in prowess, major OpenVMS users, such as Knight-Ridder MediaStream
(Philadelphia), are planning to run VMS under Windows NT. Rob Young, a
technical-support manager at Knight-Ridder, said the integration of NT management
and tool-set products into OpenVMS and Galaxies will take place gradually
over several years.
Digital's strategy is to co-opt the trend toward Windows NT by developing
leading-edge connectivity products for the Microsoft environment while continuing
to roll out proprietary products based on OpenVMS and Alpha with high-end
capabilities beyond what Windows can offer.
One Digital customer who has been following its cluster efforts closely is
Rob Young, also a technical-support manager at Knight-Ridder MediaStream.
"My interpretation is that reliability and scalability will go through the
roof [with Galaxies]," said Young. "It's a real high-end thing. You should
be able to get a huge boost just by getting Galaxies software. The
performance-improvement jump should be as high as 10 times."
"Galaxies is original," said Zalewski. "There are lots of papers on clustering,
but nothing on Galaxies. [However] one of the beauties of Galaxies is that
it won't be hard to understand."
(Next article.)
(c) 1997 CMP Media, Inc
[This article comes from EE Times in a joint cooperative effort
with the Motley Fool. For more articles like it, please look at Fool's Gold
every weekend or simply go to the Fool's Gold Mine and page through our back
issues, which all have clever and cool EE Times articles in
them.]
|