The Revenge of the Mainframe
While networked PCs get the best press coverage, Japan's "Big Three"
computer giants remain committed to offering mainframe solutions, with
CMOS technology poised as a cure for the current mainframe malaise.
by Simon Mansfield
Is the mainframe dead? Not according to Japan's computer giants, who see
huge sales to global networks taking the surviving core of the mainframe
industry into the next century. At Hitachi, Fujitsu, and NEC, system
architects are now designing the new generation of parallel-based
mainframes to meet a projected tenfold increase in data-processing demand.
Even IBM is pursuing a similar strategy, and has already begun shipping a
parallel mainframe under its System 390 umbrella.
"A three-dimensional data environment is forming from the
combination of ASCII data, sound, and image media - and, in the near
future, voice data," says Kazuo Sugiyama, manager for international
computer marketing at Hitachi Ltd. As this data flows, processing demand
will explode from the workstation onto the network. While powerful desktop
systems will be able to provide local processing for personal applications
and media data, demands on the network will grow even larger as global
databases are accessed, movies downloaded, and wages paid. Says Tetsuro
Murakami, Fujitsu's general manager for administration, international
computer systems, "data processing will continue growing. We are now in
the final stages of developing a new style of mainframe, one that is no
longer a just general-purpose computer, but a global system to serve key
operations requiring high security and reliability."
CMOS is go! By the end of 1995, Hitachi, Fujitsu, and NEC will announce
competing versions of a CPU (central processing unit) that uses CMOS
(complimentary metal-oxide semiconductor) technology rather than the
traditional bi-polar semiconductor technology used in today's mainframes.
These CMOS processors will be used in both SMP (symmetric multiprocessor)
systems - a direct descendant of traditional mainframe architecture - and
MPP (massively parallel processor) systems that use very high speed RISC
(reduced instruction set computation) processors.
The switch to CMOS technology has two advantages: the design
process for CMOS is much faster (permitting more innovation), and the
operating temperature of a CMOS processor is much lower (enabling systems
to be air-cooled). "Getting rid of the plumbing in a mainframe will cut
the cost of the average large system by around $300,000," claims Jerry
Latta, general manager for enterprise information systems at IBM Asia
Pacific.
CMOS processors will remain slower than bi-polar processors for
several more years, however. To be able to match bi-polar speeds, CMOS
will need to use 0.2-micron lithography - which will not be possible until
at least 1997. In the meantime, lower fabrication costs will allow extra
CMOS processors to be used, thereby offering a better cost-performance
ratio overall than bi-polar.
Japanese SMP systems are built around high-performance
uni-processors running in parallel. As with any multiprocessor system,
though, efficiency in scaling processors is difficult. According to IBM's
Latta. there are three key points to building an SMP system. On the
uni-processor side, there's the problem of signal delay between
processors; this requires a sophisticated switching technology that takes
account of signal latency. There is also the issue of scaleability, which
involves more than just adding an extra rack of processors. In theory,
this should double performance, but it doesn't; excellent coupling
technology is needed to maximize the performance ratio. And finally, to
make a true SMP system, the operating system must allow all the
processors and clusters to operate from one terminal as a single image.
Fujitsu is the major player in the Japanese SMP market. Fujitsu's
parallel-CMOS system, to be called the Scaleable-M series, will be
formally announced sometime in 1995. The system is to be based directly on
the M-series (first announced in 1990) as part of the Mission DC concept,
which outlined Fujitsu's vision for the future. Traditionally, Japan has
not used the term "mainframe," instead referring to these machines
as "general purpose computers." The Mission DC concept marked a
turning point for Fujitsu, with the mainframe CPU positioned as just one
of the many systems within an information network. Fujitsu has been
designing systems that take maximum advantage of a mainframe CPU for
particular applications, while spinning off other operations (such as
communications, database searching, and systems management) onto
alternative systems within the information network.
The first Scaleable-M system will feature a choice of one to eight
clusters, with each cluster comprising four CPUs mounted in a single
frame. Additional "cluster frames" can be tightly coupled in parallel,
offering a maximum of 32 CPUs. Each cluster will provide about 125
"mainframe" MIPS (million instructions per second), so that a four cluster
system can deliver 500 MIPS. "Customers can configure a system according
to their processing demands today. Later on, if they need more power, they
can just add another cluster," says Saburo Kaneda, general manager for
global server planning at Fujitsu. "By 1997, an even larger system will
be announced that offers up to 128 processors arranged in 16 clusters of
eight CPUs each." In addition to building this bigger piece of iron,
Fujitsu has continued to develop components that take operations off the
mainframe and place them on smaller dedicated systems that are able to
communicate and access data much faster.
The second Japanese player in the mainframe market is Hitachi.
After several difficult years, Hitachi is taking what some analysts see as
its last chance at remaining in the big league. Under a series of
agreements with IBM, Hitachi will use IBM's technology for both SMP
systems and massively parallel systems. "Our strategy is to work closely
with IBM in developing an SMP architecture that's compatible with IBM's
390 architecture," says Sugiyama.
IBM's Latta says the Hitachi tie-up is mainly about the cost of
developing a new generation of parallel systems. "One of the problems with
parallel systems is designing system software that can take advantage of
multiple processors. IBM believes it has solved most of these problems,
and Hitachi, in an effort to speed up its own development, will buy into
IBM's parallel CMOS technology."
However, a senior Japanese computer executive (who asked not to be
named) cautions that Hitachi faces many serious problems. "It is doubtful
that the company, over the long term, can remain an independent vendor.
Although most Japanese customers regard Hitachi's after-sales support as
very good, its only real chance for survival long term will be as a
systems integrator. And this will depend on the approach taken by
Hitachi's management in restructuring as an open systems company."
The third major Japanese player in the SMP game is NEC. Although
largely unknown among international mainframe customers, NEC holds 17.7%
of the mainframe market in Japan - this in addition to its almost 50%
share of the Japanese PC market. Furthermore, NEC has a major tie-up with
Europe's Bull Computers for the development of CMOS processors. "Data is
becoming very complex and very large. In addition, network speeds are
rising very quickly. To meet these demands, the processing has to be done
much faster," says Hiroo Shin, assistant general manager for large system
planning at NEC. "Although most of our emphasis is on PCs publicly, NEC
has major programs underway to develop both SMP systems and massively
parallel systems using CMOS technology supplied by Bull." Japanese MPP
systems Another key area for hardware is massively parallel processing, or
MPP for short. Most of this work has taken place in the supercomputer
world, where computation-intensive applications are required. But a
growing number of vendors in the US and Japan are applying this technology
to develop a new class of systems that will bring the power of
supercomputer technology into the commercial mainframe world. Opinion is
divided on how soon this technology will be commercialized.
Initially, the industry began with a shaky start, following the
failure of two US-based pioneers in the area (Kendall Square Research and
Thinking Machines). Japan's big three, along with IBM and several others,
however, have continued their research. Hitachi and NEC are the most
aggressive in terms of hardware, with both companies committing to a
processor architecture. "For high-end database servers, we will adopt
IBM's Power Parallel processor - a second generation RS6000 processor that
allows systems to be configured with thousands of parallel processors,"
reveals Hitachi's Sugiyama. These systems will be used for managing large
database systems, where thousands of requests flow in from the network.
The vendors hope that MPP systems will find wide application among OLTP
(online transaction processing) users. Considerable effort on the software
side will be required before a large number of sales are realized,
however.
Fujitsu remains relatively cautious on committing to a processor
architecture, saying that the system software has a long way to go before
it becomes a serious commercial system. Nonetheless, Fujitsu has a number
of maspar (massively parallel) programs under development that will
eventually flow into the commercial market. The company has set up three
parallel software research labs (in Japan, Australia, and the UK) to
develop maspar system software, and it has developed a 64-bit SPARC
processor that could lay the foundations for a maspar CPU. In the US,
meanwhile, a 100% Fujitsu-owned subsidiary, HAL Computers, is developing
high performance RISC boxes that operate in parallel. Market demands
Richard Sullivan, partner in charge of technology integration for Andersen
Consulting in the Asia-Pacific, believes the issues facing companies today
go beyond processing power. "Given the restrictions of traditional
mainframe applications, the benefits of parallel systems will be
restricted to a handful of applications, such as online operations, unless
major modifications are made to the software, from the ground up," he
says. "For example, batch processing is a monolithic task difficult to
break into smaller pieces that can then run on parallel processors. Batch
speeds have essentially topped out. Additional processors will only serve
to process higher volumes, with little impact on the speed each job is
processed at."
Hardware development remains way ahead of the software - both
system and application software. "And there is no magic bullet, such as
middleware, to solve this problem," says Sullivan. "Eventually, vendors
and customers will have to redesign the application software and the way
it operates on the system. The shift to parallel will accelerate the move
from batch to interactive processing; in practical terms, this is already
a clear trend. The message for MIS managers is two-fold: Migrate
applications to online while reducing batch operations, and look very
hard at the different system software to clearly establish what offers the
best performance of the given technology."
The ultimate success of parallel processing will depend upon the
design of highly complex databases and system software that help
applications run effectively on parallel processors. Up front, a network
of PCs can look cheap. But when you start to add in
network cards, the cost of new software, and support, the costs go up.
Furthermore, security and reliability issues will dictate what is the most
appropriate IT (information technology) solution for an organization.
"Customers are starting to understand that the total cost of ownership for
an open systems solution is not always cheaper than a large mainframe
solution using proprietary systems," notes Fujitsu's Murakami.
"Furthermore, they realize PC-based solutions cannot meet the processing
demands of many large enterprise applications like OLTP, batch
processing, and large database management systems - applications that are
critical to many organizations."
Murakami suggests modes of transportation as an apt analogy. "We
all know there are trains, planes, and automobiles. But you don't compare
a motorcycle to a truck, or a bullet train to a family car. The only
common aspect is transport." Computers are the same, he says. PCs and
mainframes can both crunch data, but their speed and power differ between
platforms and applications.
In this era of downsizing, networked PCs have taken the spotlight.
But if Japan's big three mainframe makers have their way, the mainframe
may be poised for a comeback.
