Viewpoints of history can help us assess the present and the future. Suresh Shukla and I exchanged some notes in this regard. Over the last thirty-plus years, HPC has undergone multiple metamorphoses, but fewer and fewer people that worked on fast computers in the early '70s are actively working today with the latest HPC equipment. We thought it worthwhile to start documenting their viewpoints about HPC as it was. Shukla kindly agreed to start the series with his viewpoints. This is part II of his two-part series.
Shukla worked in HPC as it is applied within industry. His viewpoints may be different from those of people who worked in a government lab or at a university. These are Shukla's personal views. We look forward to publishing viewpoints of others who have spent many years in the field of HPC. We thank independent consultant Steve Conway, Earl Joseph from IDC, and Suzy Tichenor from US Council on Competitiveness for reviewing these articles.
– Tim Curns, Editor
THE SMALL FISH EAT THE BIG FISH (1987 – 20??)
We looked at the earlier period (1972 – 1987) in part I of this series. I suggested then that the history of that period can help us strengthen some actions that are being taken today. Programs like HECRTF are identifying grand challenge applications that would benefit from major advances in high-end computing. Support for developing and validating critical applications needs to be understood and strengthened. An important initiative in this regard is the DOE's INCITE program, wherein Dr. Orbach is making high-end computers available to industry to develop and test new applications for solving grand challenge problems. Industry needs to take full advantage of programs like this.
The period from 1987 onwards differentiates itself from the previous one by the increased attention given to the affordability of HPC. Many events happened before 1987 that led to this shift in attention. Moore's law, first stated in 1965, notably proved itself in the period through 1987 by shrinking the circuitry on a chip, increasing its power and reducing its cost by many orders of magnitude. IBM entered the PC business in 1981, validating the concept of affordable desktop computing. Single-CPU performance already started pushing against the limits of physical science, bringing increased attention to parallelization as the major technique for speeding up computations. HPC in this period started growing more in breadth rather than in depth.
Not only did affordability pervade the HPC segment, it pervaded the whole computing industry, The relative cost of HPC vector platforms started looking much larger in comparison to the cost of business computing. The computing industry responded, and in the late '80s, 32 mini-supercomputing companies such as Convex and SCS sprang up to address the appetite for lower-cost HPC computing.
This also affected the R&D efforts. As affordability became the watchword of the day, budgets for application development, validation, and improvement were cut to a minimum. In the industry sector, they became almost non-existent. Unlike in the '60s, very few new applications were developed, especially for harnessing the power of capability platforms. Most of the R&D activity with respect to applications was confined to porting them to lower-and-lower-cost platforms.
With minor exceptions, we in the HPC community have spent the last 20 years without developing any applications that address contemporary grand challenge problems. Efforts to move to even lower-cost platforms became so successful by mid '90s that all the 32 mini-super computing companies simply vanished or were absorbed into larger computing companies. With broad commercial access to the Internet in 1992, and with the ASCI contract coming into being, clusters (and to a certain extent, grids) became the machines of choice for most applications, and so they remain today. Many of us adopted a poor metric to define requirements, i.e., theoretical GFLOPs, unaware that it would lead suppliers not to worry about actual delivered performance as much.
It is definitely a sign of progress that thousands of analyses can be performed on clusters routinely today that were the private domain of a select few supercomputers before 1987; but we are not solving any new grand challenge problems. The current problems we solve can also be solved in third world countries on clusters.
No wonder during this same period very few have “taken a fancy” to HPC analyses. Again with rare exceptions, HPC doesn't entice young people to solve new grand challenge problems. Most of today's “grand challenges” have migrated toward the World Wide Web. We see more talent recruited into pure computing disciplines (computer science) and less into HPC applications (computational science). This has created a downward spiral of talent and innovation in HPC. We in the USA are becoming less innovative and less competitive compared with the rest of the world, as far as HPC is concerned.
Looking at the revenue statistics of HPC hardware suppliers, as gathered and projected by IDC, only clusters and grid computing seem to have the momentum needed to sustain themselves. The same doesn't appear to be true with suppliers of high-end computers. Until applications are developed that will once again keep lots of high-end machines busy and profitably deployed in industry, their development may need continued support from government and universities.
CAN BIG FISH AND SMALL FISH CO-EXIST?
This is definitely possible. In the past five years, government agencies such as DARPA, DOE, NASA and NSF, along with ITRD and Congress, are realizing the value of such a co-existence for national competitive advantage. HECRTF and other initiatives are also attempting to revitalize high-end computing.
Industry, though, is lagging far behind in this effort. It hasn't been able to shift its attention from affordability to the value of solving grand challenge problems yet, It needs to cooperate with application development and validation efforts, such as the one proposed by DOE through their INCITE program. This program is inviting industry to submit proposals to utilize the DOE's high performance computers to solve its most challenging, and therefore most competitively important problems.
If industry doesn't focus on these high end problems, tomorrow's crucial high-end applications would likely be developed in some other country.
At the HPC User Conference, Boeing recently identified many grand challenge problems for aerospace in the CFD area that would need at least an order-of- magnitude more computing power than what is available today. If industry identifies more applications as candidates for development and validation on more-capable high-end computing equipment, I will consider my efforts in writing this two-part article worthwhile.
In 2004, Mr. Shukla was an HPCwire “Top Person to Watch,” and he is an executive member of the HPC User Forum committee. He can be emailed at [email protected].