Dealing with Client/Server
Issues in Purchasing and Implementation
When More CPUs Are Better
Moving to parallel processing could be the right step for your IS department,
but perform the necessary evaluation before you leap ahead.
By Sally Atkins
As a member of a design or development team, when you sense that a new technology
may be more appropriate than the mainframe or a single server, it can be
intimidating to face an IS department and its advisors who are traditionalists.
I found myself in such a predicament a few years ago in recommending parallel
processing. Today, that would be easier. Buying multiprocessor systems is
no longer for the early adopters alone. Gartner Group recently estimated
that this $1 billion market will grow 50 percent annually.
For the record, let's clarify usage regarding parallel processing and its
more specific incarnations. Parallel processing is the umbrella term
for a computing technique in which many operations are performed simultaneously,
using many microprocessors. A "coarse-grain" system, often identified
as symmetric multiprocessing (SMP), contains relatively few but powerful
processors. A "fine-grain," or massively parallel processing
(MPP), machine contains up to thousands of smaller ones.
Increasing name recognition and accelerating market share may open the door
to considering an idea in your company, but they are not a sufficient reason
to recommend the purchase of parallel processors. Applying more generic
business justifications to your project is a better way to help determine
whether you should take a closer look at parallel processing. Here are some
useful considerations. (For a full discussion of this technology applied
to data access, see The
Race for Data Access)
No Pat Answers
In price/performance ratio, parallelism far outstrips mainframes and single
RISC or Intel servers for many applications. Actual numbers will vary depending
upon the application.
It can be reassuring that early adopters have gone before you. Multiprocessor
systems have been used successfully for online transaction processing, data
warehousing, decision support, simulation and financial modeling. Vendors
can provide sample success stories to help with your cost/benefit analysis.
Database vendors finally have arrived on these platforms. Informix took
three years to rewrite its core database architecture to provide clustered
SMP and another year to accommodate MPP functionality. Sybase, Oracle and
IBM now also have integrated parallel processing capability into their database
Scalability is a major benefit. SMP is available from the low end, with
two- and four-way machines, up to the high end, where systems with eight
or more processors are outperforming mainframes for many applications. When
anticipated correctly, upgrades are relatively straightforward.
SMP cluster systems (multiple processors on multiple linked machines)
can overcome some of the operating system limitations in handling SMP. Hewlett-Packard,
for example, is touting the advantage of parallel server technology over
more traditional MPP architectures.
There are multiple sources. All major hardware vendors have a multiprocessor
product line, so you can shop around and count on continued competition,
variety and innovation among major systems.
Choosing SMP, clustering or MPP should depend upon your specific applications
and goals. Solutions change almost as fast as definitions. It used to be
that data warehousing, which must store large volumes of data and provide
reasonable response times on queries for a large number of users, was an
application for MPP architectures. When detailed queries on massive amounts
of operational data were required in realtime--for instance, if you were
Mattel and needed to know how many Busy Gal Barbies sold in Boston in the
past hour--MPP was the answer. Now, SMP solutions and clusters should be
SMP is sure to become more popular as more tools and applications are developed
for these environments. And clusters do not necessarily require new hardware,
while MPP systems usually do.
Stop, Look and Listen
It is best to have an experienced, vendor-neutral guide when going through
this the first time. A few years ago, I was helping an internal IS team
responsible for recommending a server for a huge financial modeling, or
number-crunching, application. The IS department had initially recommended
adopting Windows NT on an Intel processor.
Once we modeled the application conceptually and simulated it on the users'
mainframe, we found single-processor server architectures to be inappropriate.
The application took 24 hours to run! There was not enough mainframe capacity
in this $40 billion corporation to handle such a new application. We ended
up testing with an eight-way Sun Sparcserver. The simulations, not possible
before the new financial modeling had been done, ran in a quarter of the
time they would have taken on the mainframe. Seeing the proof of the simulations
left no business choice but to adopt the new parallel technology.
We took several lessons from this experience. One was to slow down the users
and their eager programmers long enough to do simulations before entering
the recommendation stage. The full model allowed us to understand the nature
of the processes and data we were dealing with, and we gained the will to
go ahead and build applications that were not feasible with single-processor
We also learned to use the knowledge bases of the hardware, operating system
and database vendors in the presales process, just as we'd consult with
any other member of the team. Some vendor alliances are stronger than others.
We found that engineering partnerships are a lot deeper and more valuable
to users than mere marketing alliances. These show up in presales modeling
efforts and can help you choose vendor partners that are capable of working
Moving to parallel processing, especially in a distributed client/server
environment, is a big step. It may not be appropriate for every need, but
today there is no good reason not to consider it as an alternative.
Sally Atkins is president of IST Consulting, an affiliate
of NetSource, Inc., based in Boston. She can be reached at Sally@kins.com.