Last week, I was reading this press release from a company claiming it has developed the first mathematical derivation of Amdahl's Law. The Colorado-based company, Massively Parallel Technologies, Inc. (MPT), doesn't offer a paper about this claim on its web site. Instead, it says its software "makes supercomputing faster and cheaper than ever before." MPT adds that any cluster using its software can achieve almost 100% efficiency. For example, they say they operate a 255-processor cluster "that accelerates difficult problems involving heavy cross-communication by an astounding 250 to 350 times" or that a "cluster with 1023 nodes requires over 40,000 times less communication time steps than its conventional counterpart." The company looks real, but its claim seem extravagant. Has anyone used this technology? Is there something real behind these claims? Read more and post your comments.
Let's start with two paragraphs from the gibberish MPT press release
The Massively Parallel approach, now confirmed by Amdahl's law as the most efficient method of parallel processing, concentrates on improving the efficiency of the communication methodology between processors. This has resulted in systems that exceed 90 percent efficiency rather than the previously believed limit of 20 percent. The breakthrough also allows systems to be built with low cost commodity hardware to solve problems in minutes rather than hours.
Impressed with MPT's achievements, Dr. Amdahl now sits on the company's board of advisors. "It is a great pleasure for me to see the fruits of my work being realized so dramatically by MPT," said Amdahl. "MPT's technology will allow supercomputing power to be used by anyone, anywhere and I have no doubt it will result in dramatic achievement and scientific advancements. The sky's the limit."
OK, let's switch from the press release to what the company says about its technology.
One of the opening paragraph sounds reasonable.
There are three primary data movement times which are important to parallel processing. They are data input, data output, and cross-communication. HOWARD [,their trademarked software,] has several novel solutions for I/O. One solution balances the time it takes to "cascade" a problem called 'ψ’ with the amount of time it takes to perform the I/O of that problem '# time-steps'. HOWARD does this via the use of multiple communication channels 'ψbase' and MPT’s proprietary "Howard Cascade."
But later down through the page, things start to appear as unreal. MPT says for instance that "in order for a conventional system with 41,391 nodes relying upon fast connectivity to equal the performance of a HOWARD system using office LAN quality connectivity (100Mb/s), it would need channel speeds of 15 Tb/s."
Socialism proposes no adequate substitute for the motive of enlightened selfishness that to-day is at the basis of all human labor and effort, enterprise and new activity.
—William Howard Taft (1857–1930)
By the way, do you know many conventional systems with more than 40,000 nodes exist today? None -- if we except the soon-to-be-finished IBM's Blue Gene/L system.
It's time to switch to the MPT's applications page, where the company explains its secret.
The secret is in MPT’s revolutionary new approach to parallel processing. HOWARD, unlike conventional architectures, does not rely upon fast low latency communication channels to connect processors in order to achieve the parallelism necessary to dramatically boost processing performance. Instead HOWARD is based upon an entirely new mathematical geometry in the distribution of problems to multiple processors and for cross-communicating partial solutions from one processor to another. The geometry utilized in a HOWARD cluster of 1023 nodes, for example, can utilize office LAN quality 100 Mb/s Ethernet connections for its channels and yet outperform conventional clusters with expensive 10 GB/s connectivity. In fact, for the most difficult of all cross-communication exercises, an "all to all" exchange between processors, the geometry of a HOWARD" cluster with 1023 nodes requires over 40,000 times less communication time steps than its conventional counterpart.
And MPT gives an example about the oil and gas industry, a sector I worked with for a number of years.
HOWARD is already revolutionizing the oil and gas industry. MPT currently operates a 255-processor cluster that processes seismic data so rapidly that it can produce nearly five million dollars worth of finished seismic data per year at an annual operating cost of just $250,000.
And if these fabulous numbers haven't convinced you yet, what about the costs of an MPT solution?
The price per processor in a conventional parallel processing system is usually in the $10,000 to $100,000 range, while the efficiency of the system is usually under 10%. Thus a solution on a conventional system costs 40 to 400 times the price that it would cost if delivered on a single processor workstation in which the total cost of the system is around $2,500.
How is it possible that such a company has been ignored by the media -- if its claims are true?
It is not surprising that the motion picture relies to a considerable extent on devices borrowed from the theatre. There are many superficial points of resemblance between the two arts. Films are shown in auditoriums which do not differ in any marked degree from other playhouses. Stage actors perform in films. Stage training is still regarded as a fairly satisfactory prerequisite for appearance before the camera. Actors, directors, and writers move from stage to screen with comparative ease and with what seems to be minor adjustments of their techniques and methods. It is often customary to draw a stage curtain back and forth to mark the beginning and end of the drama projected on the screen within the proscenium arch.
—John Howard Lawson (1895–1977)
By the way, they really filed an application for a trademark for HOWARD. You can check with the United States Patent and Trademark Office. MPT files its "computer software in the field of parallel processing" under number 78513638 on November 9, 2004. But a trademark doesn't imply that's a product is good.
I've been involved for twenty years in supercomputing, working with many different parallel systems. And I never saw something similar to MPT's claims. Is the company on something real? I've serious doubts, even if MPT received some grants from DARPA. If you ever used their software, or heard about it, please post your comments.
Sources: Massively Parallel Technologies press release, January 18, 2005; and MPT website
Related stories can be found in the following categories.
Computers
IT
Software
Supercomputers.