Market researcher Palo Alto Management Group believes the massively parallel processing market could be worth at least $10,000m by the year 2000, hitting $85,000m by 1998 on the way. Two-thirds of the market, it estimates, will be for commercial transaction processing or decision support systems, with old-style mainframe-type pricing and profit models being quickly superseded. It reckons, surprisingly, that Tandem Computers Inc has 35% of the market currently, AT&T Global Information Solutions 29%, Intel Corp and Thinking Machines Corp 6% each and IBM Corp, MasPar Computer Corp, Meiko Scientific Co and nCube Corp 2% each. The remaining 15% is split between some 20 or so other vendors, the group says. Although Tandem doesn’t necessarily leap to mind when considering massively parallel processing, sales of clustered fault-tolerant Cyclone systems and other machines with multi-hundreds of CPUs generated $525m of revenue for the Cupertino, California company last year, representing around a quarter of its turnover.
Tandem’s original stock certificates bear a description of the company as a provider of parallel computing systems. Like names given to some other kinds of information technologies, massively parallel processing is something of a misnomer. It has many different interpretations and parallelism can be achieved in a variety of ways. Indeed, using a straight definition of massively parallel processing architectures that string hundreds or thousands of CPUs together such as Cray Research Inc, nCube, Thinking Machines or Parsytec, excludes clustering – used by Tandem – or other techniques employed in the successful AT&T Teradata systems. In a limited definition, the market was probably only worth only around $275m last year, Palo Alto Management believes. It has expanded its definition so that most systems that have at least 64 CPU nodes are now counted, including workstations or servers linked by clustering software. Think in terms of scalable computing, beyond shared memory symmetric multiprocessing, says Palo Alto Management’s Michael Burwen, the study’s director. Massively parallel processing systems are, in his estimation, the only machines that can usefully scale memory and disk size to the Terabyte level. In the expanded context, it estimates the 1994 market will be worth some $1,500m. There are various mechanisms for stringing processors together in a single, or multiple machines – and they are hotly disputed by the competing firms – but with the increasing emphasis on commercial installations, clustering techniques, which enable users to add nodes according to requirements, are likely to predominate, Burwen believes.
By William Fellows
Firms such as IBM and Convex Computer Corp have an additional advantage in that their massively parallel processing systems enable users to run existing software in uniprocessor mode. If IBM’s parallel gambits, the Powerparallel SP1 and SP2 take off, the company could end up with up to 25% of the 1998 market, Burwen estimates. How quickly, and how widely, the technology is adopted depends upon getting robust, secure, high availability middleware – mostly relational databases – up on massively parallel processing systems for commercial customers. And it’s a real problem at the moment, Burwen says. Although Oracle Corp, Informix Corp, Sybase Inc and IBM all have parallelised versions of their databases, and there are implementations for various architectures, there aren’t yet any real sites doing real mission-critical commercial processing – the implementations simply aren’t up to the task yet. As such, there are not any meaningful performance figures or comparisons. Indeed, Computerworld reports that performance shortfalls at its one test site – New York insurance firm Kwasha Lipton, which is using Navigation Server on an NCR 3600 parallel system – have forced Sybase to postpone shipping the parallel version of its System 10 database for six months. Navigation Server was announced along with System 10 last May and was to have shipped this summer and is being dev
eloped upon the NCR 3600. Implementations for other systems, including Sun Microsystems Inc, Hewlett-Packard Co and IBM are due, although AT&T Global will have it for six months before rivals as part of their agreement.
Navigation Server is an add-on module – Informix and Oracle will parallelise at the core level. With six more companies due to come to market with massively parallel processing systems over the next year or so, a major shake-out is likely, Palo Alto Management believes, and it reckons that the biggest challenge for massively parallel processing vendors over the next five years will be to figure out how to make money out of the things. Players in the massively parallel processing market include eight firms with annual revenues above $25m, 16 with less, and six only now coming to market. The market clearly can’t sustain 30 vendors right now. New massively parallel processing-class offerings due this year include those from Jacques Stern’s Advanced Computer Research International SA, Unisys Corp with Intel, Hitachi Ltd, Tera Inc, Pyramid Technology Corp and although given its recent thoughts on the rush to massively parallel processing it probably won’t be pleased with its inclusion here – Sequent Computer Systems Inc, which has just upgraded its Symmetry SMP series to the 5000, with a new 240Mbps Highly Scalable Bus and clustering options. The price-performance characteristics of massively parallel processors seem to insist that the architecture becomes tomorrow’s equivalent to the mainframe of today, but how many of the current players will live to see their day in the sunshine is decidedly moot.