From the early days of the industry, supercomputers have been pushing the boundaries of IT, identifying the key barriers to overcome and experimenting with technologies and architectures that are then incorporated into the overall IT market a few years later. While we generally focus on their computational capabilities as measured in FLOPS, - Floating-point Operations Per Second, - supercomputers have been at the leading edge in a number of additional dimensions, including the storage and analysis of massive amounts of data; very high bandwidth networks; and highly realistic visualizations.
Through the 1960s, 1970s and 1980s, the fastest supercomputers were based on highly specialized, powerful technologies. But, by the late 1980s, these complex and expensive technologies ran out of gas and parallel computing became the only realistic alternative to scaling up performance.
Instead of building machines with a small number of very fast and expensive processors, the early parallel supercomputers ganged together 10s, 100s, and over time 1000s of much less powerful and inexpensive CMOS microprocessors, similar to the micros used in the rapidly growing personal computer and workstation industry. A similar evolution to microprocessor components and parallel architectures took place a few years later in the mainframes used in commercial applications.
The transition to parallel supercomputing was seismic in nature. Everything changed, from the underlying computer architecture, to the operating systems, programming tools, mathematical methods and applications. It took considerable research and experimentation to learn to effectively use these new kinds of machines. Moreover, there were widely different parallel architecture designs, some coming from universities and others from industry. It wasn’t clear at all which designs worked well for different kinds of applications and would thus be commercially viable.
The Department of Energy (DOE) national labs have long been among the world’s leading users of advanced supercomputers and played a leading role in the transition to parallel architectures. In 1983, the DOE’s Argonne National Lab established the Advanced Computing Research Facility (ACRF), an experimental parallel computing lab which brought together computer scientists, applied mathematicians and supercomputer users and vendors to learn how to best use this new generation of parallel machines.
This past May, Argonne convened a Symposium to mark the 30th anniversary of the ACRF. The Symposium looked both at the progress made in parallel computing over the past 30 years and the major trends for the future. I attended the Symposium and led a panel on The Impact of Parallel Computing on the World.
At the time, professor Roothaan was consulting with IBM on the design of of a new generation of supercomputers, and I got involved looking at how to best program mathematical algorithms for these machines. A few of the IBM people I worked with encouraged me to apply for a position in computer sciences at IBM’s research labs. Switching from physics to computers sciences was not an easy decision. But I finally realized that I enjoyed the computing more than the physics, and in June of 1970 I joined the computer sciences department at IBM’s Watson Research Center, the beginning of my long career with the company.
IBM stopped developing supercomputers in 1969 for a number of business reasons. But, by the mid-1980s it was becoming clear that by not working on such leading-edge products, IBM was cutting itself off from many of the IT innovations coming out of research labs and universities. The decision was then made to return to the supercomputing business, initially through extensions to mainframes in the late 1980s, a relatively modest effort which I led for a few years. The resulting machines were not all that competitive with the leaders in the field, but got IBM back in the game.
In the early 1990s, I was general manager of a much more successful IBM supercomputing effort, the Scalable POWERParallel (SP) family of parallel supercomputers which were based on IBM’s POWER microprocessors. Argonne was a close partner of IBM in developing the original SP machine. Argonne acquired one of the first supercomputers we shipped in 1993, and given their singular expertise with parallel architectures, they gave us invaluable feedback and assistance in its continuing development.
IBM has since become a leader in supercomputing, with 160 of the top 500 systems in the latest Top500 list and 5 of the top 10. The close working relationship between IBM and Argonne has continued over the past 20 years. On July 1, Argonne officially dedicated Mira, which at almost 10 petaflops (10 followed by 15 zeros) is the fifth most powerful supercomputer in the world and will be used in scientific research ranging from the action of blood vessels to the origins of the universe.
The performance advances of supercomputers in these past decades have been remarkable. The machines I used as a student in the 1960s probably had a peak performance of a few million calculations per second. Gigaflops (billions) peak speeds were achieved in 1985, teraflops (trillions) in 1997, and petaflops in 2008. According to the latest Top500 list, the fastest machine, China’s Tianhe supercomputer, can now achieve over 30 petaflops. The supercomputing community is now aiming at exascale computing, - 1,000,000,000,000,000,000 calculations per second, - which is expected to be achieved sometime by 2020.
But, once more, the current approach to supercomputing is running out of gas. On May 22, the House Science Committee held hearing on America's Next Generation Supercomputer: The Exascale Challenge. Rick Stevens, - Argonne’s Associate Lab Director responsible for Computing, Environment, and Life Sciences research and professor of computer sciences at the University of Chicago, - was one of the witnesses at the House hearing. In his testimony, professor Stevens listed the “five major hurdles that must be overcome if we are to achieve our goal of pushing the computing performance frontier to the Exascale by the end of the decade:
- We must reduce power consumption by at least a factor of 50.
- We must increase the parallelism of our applications software and operating
systems by at least a factor of 1,000.
- We must develop new programming methods to increase dramatically the
number of programmers that can develop parallel programs.
- We must improve memory performance and cost by a factor of 100.
- We must improve systems reliability by at least a factor of 10.”
The 50X reduction in power consumption feels like the most challenging hurdle. Without such improvements, the energy cost of operating an exascale system would be prohibitive - $100s of millions per year. Industry knows how to achieve a 5X reduction, but to achieve the necessary 50X just about all components must be redesigned, including the processors, memory, storage and overall wiring and interconnections.
Achieving such breakthrough reductions in power consumption will have a major impact across the IT industry, from the billions of mobile devices and trillions of smart sensors being deployed around the world, to the huge cloud-based data centers being built to support all these devices as well as to store and analyze the massive amounts of data collected every day. “Given that an estimated 5 percent of global energy consumption, and of global carbon dioxide emissions, is attributed to computing services, energy efficiency improvements of this magnitude could have a significant impact on the environment,” adds Stevens.
Over the past few years, the Department of Energy held a number of workshops bringing together more than 1,200 scientists and engineers to identify some of the key problems whose solution require exascale-class computing. In his testimony, Stevens mentioned some of these problems:
- “use first principles to design new materials that will enable a 500-mile electric car battery pack;
- build end-to-end simulations of advanced nuclear reactors that are modular, safe and affordable;
- add full atmospheric chemistry and microbial processes to climate models;
- increase the resolution of climate models to provide detailed regional impacts;
- model controls for an electric grid that has 30 percent renewable generation;
- create personalized medicines that will incorporate an individual’s genetic information into a specific, customized plan for prevention or treatment;
- study dark matter and dark energy by building high- resolution cosmological simulations to interpret next generation observations.”
In the decades ahead, Argonne and the rest of the supercomputing community will keep pushing the IT boundaries, enabling us to better address the most challenging 21st century scientific, economic and societal problems. It will surely be as exciting a journey as has been the 30 years of parallel computing at Argonne we just celebrated.