SGI
and Intel
It's
uncertain what will ultimately happen to supercomputer vendors with
the commoditization of supercomputing that Computing Fabrics will
bring. With the advent of PCs, many makers of minicomputers went
the way of the dodo. Other midtier systems vendors, such as Sequent
Computer Systems Inc., managed to survive the onslaught of PC-based
servers and the commoditization of the minicomputer market by exploiting
commodity processors in their parallel designs.
Silicon
Graphics Inc., by planning to offer systems that scale from 1 to
more than 1,000 Intel Corp. Merced processors, is following an even
stronger strategy. Its alliances with Intel and Microsoft Corp.
should enable SGI to not only survive, but to thrive.
Besides
this, vendors such as SGI/Cray possess far-reaching expertise that
companies like Dell Computer Corp. and Compaq Computer Corp. can't
approach. Challenging computational problems remain unsolved, for
lack of sufficient horsepower. The Department of Energy's supercluster
at Los Alamos National Labs, which unites 48 SGI/Cray Origin 2000s
of 128 processors each, is just the beginning.
Computing
Fabrics presage the era of Petaflops computing, where systems will
deliver one million billion floating-point operations per second,
roughly 10 times the total networked processing power in the United
States just 3 years ago.
Possessing
the expertise for these grand-scale projects means a continued existence
for supercomputer vendors in general and SGI in particular, but
in the new era of Computing Fabrics, these vendors may begin to
look more like super-system integrators and pioneers of enabling
technologies, and less like system vendors.
To
Intel, the early stages of Computing Fabrics will look like business
as usual, with new markets and an expanding customer base. Intel's
core business will be strengthened, because its general-purpose
processors better exploit fabrics than do dedicated application-specific
integrated circuits.
The
same logic reinforces Microsoft Corp.'s position that a general-purpose
operating system like Windows CE makes for more flexible embedded
systems. As fabrics take hold and buses converge with networks,
Intel might take steps to better exploit potential synergies between
its system board, networking and microprocessor divisions. Just
as it purchased Corollary for its 8-way Perfusion architecture,
in time, Intel could acquire the networking and systems groups of
SGI for its Computing Fabrics technologies.
Microsoft
Microsoft
has quietly been conducting research in distributed operating systems,
addressing the hard problems of distributed processing including
support of a single Windows object space across numerous processors.
This has yet to transform NT into an enterprise-class operating
system, however, and no one, including Microsoft, is holding their
breath.
As
Intel-based Fabrics are announced, the pressure of the marketplace
is sure to encourage Microsoft to acquire the requisite technology
from a third party. SGI's Cellular Irix is a likely candidate, and
Microsoft and SGI are in discussion on the possibility.
Sun
Microsystems
What
of Sun Microsystems, which remains committed to its noncommodity
microprocessors? Although Sun has the distributed software infrastructure
for a Computing Fabrics future, its hardware direction is not so
well-adapted.
On
the hardware side, Sun is trying to go it alone, with a processor
architecture to be deployed from embedded systems up to super servers.
However, Sun won't benefit from the economies of scale that those
using Intel processors achieve, notably SGI and HP. Furthermore,
Sun has neither a scalable interconnect story--it's using a single
big, fast crossbar switch--nor a single system image that extends
beyond 64 processors, due to Sun's use of SMP architecture.
It
will be hard for Sun to compete against SGI's high-volume Intel
processor, scalable architecture and modular Fabric. Increasingly,
Sun will be pressured to either "go Intel" or partner with IBM and
share its fabs, because SGI and others offer major competition in
the enterprise space of database servers, data mining, decision
support systems and Web/intranet hosting.
On
the software side, however, Sun may well be holding four aces in
Java, Jini, JAIN (Java Advanced Intelligent Networks) and JavaSpaces--distributed
technologies for which Computing Fabrics are likely to stimulate
great demand. Sun could well put together a Java-based competitor
to Microsoft's Millennium, using Enterprise Java Beans for code
mobility and CORBA (Common Object Request Broker Architecture) and
Internet Inter-ORB Protocol for distribution transparency.
IBM
The
investment in legacy software that only runs on the mainframe will
not last forever. The power, cost savings, and ease of administration
of Computing Fabrics will overshadow even the expense of redeveloping
legacy applications.
Can
IBM move into the Fabric space? It certainly possesses expertise
in distributed systems, having debuted n-tier architectures in the
early '80s, updated MVS systems to Parallel Sysplex, and invested
R&D funds in projects such as T Spaces. It's more a question
of management and marketing, and if IBM's move into Java is any
indication, it may well be able to compete in the Fabric space,
especially if it partners even more extensively with Sun. Make no
mistake, however: This will be the IBM mainframe's final battlefield,
up against the combined talents of Intel, Microsoft and SGI/Cray.
Network
vendors
Vendors
of network hardware and software are likely to enter the Computing
Fabrics fold as well. Switches are likely to evolve to directly
support the rich topologies of Fabrics as well as their extreme
performance requirement for very low latencies and cache coherence.
Network
operating systems may evolve into cellular operating systems. Taken
together, this means that network vendors have an excellent opportunity
to challenge systems vendors in the midsize to large systems market,
a radical shift.
Software
development
The
ramifications of Computing Fabrics for applications and system software
are significant. Current applications will be compatible with fabrics,
but software will need to evolve to exploit the fluid system boundaries
of fabrics.
DBMSes
are likely to lead the way here, based on their utilization of both
in-box parallelism and multisystem clustering. But even DBMSes will
need to be further "exploded" into interoperable components that
can more fully utilize the fabric.
To
fully exploit fabrics and their fluid system boundaries, a rapprochement
is needed between these vastly different programming formalisms.
Microsoft's Millennium points to a partial solution, taking code
developed along object-oriented lines and automatically distributing
its components using a distributed object infrastructure running
atop massively distributed clusters.
Logical-level
models and views become increasingly important as data and processing
are increasingly and dynamically distributed over the Fabric, in
much the same way as the relational model satisfied the need for
a unified logical view of data across disk farms. Such logical models
will by necessity need to unify a wide variety of underlying systems,
including COM, CORBA, and Java objects, relational and object DBMSes,
Extensible Markup Language and HTML documents, multiple file systems,
and varying directory structures.
Computing
Fabrics are also likely to encourage, if not require, migration
and adaptation of solutions across system levels. For example, distributed
DBMS techniques may need to be adapted by cellular operating systems
to maintain cache coherency across loosely coupled systems.
Erick
Von Schweber is Chief Science Officer for Infomaniacs, a think tank
in Sedona, Ariz specializing in technology convergence.. Linda Von
Schweber is Chief Creative Officer for Infomaniacs. They can be
contacted at thinktank@infomaniacs.com
or www.infomaniacs.com.
|
