Since 1992 there has been a moratorium on nuclear testing in the United States. This has forced the National Nuclear Security Administration (NNSA) to develop a program to maintain our nuclear stockpile by other means. The Advanced Simulation and Computing Program (formerly known as the Accelerated Strategic Computing Initiative, or ASCI) was established to create this capability. Under ASC, computer simulation capabilities are developed to analyze and predict the performance, safety, and reliability of nuclear weapons and to certify their functionality.
HPCwire got the opportunity to speak with the director of ACS, Dimitri Kusnezov, about the nature of the program and the significance of the IBM Blue Gene/L (BG/L) supercomputing technology in maintaining our nuclear stockpile.
HPCwire: Can you help us understand the big picture for NNSA's ASC program?
Kusnezov: Sure, it is actually a rather simple concept. If you don't want to test nuclear weapons like we did in the past on a weekly or monthly basis, but you still want to maintain our country's nuclear competence, you must develop the means to do it virtually. That in a nutshell is the ASC program. We push the limits of computational science and supercomputing to meet our evermore challenging national nuclear security needs.
We do so much virtually today, that computer simulation has become second nature — similar to email's importance to how you and I work today. Today, NNSA uses the Stockpile Stewardship program to ensure the safety, security and reliability of the nation's nuclear deterrent without testing. It pulls together the theoretical models, experimental results, and legacy databases into an aggregated picture that weapon scientists use to assess and gain confidence in the viability of the nation's nuclear stockpile or how to deal with other emerging needs. Without this simulation capability, the country would have to consider renewed testing, which is something we don't want to do.
The world has changed considerably since the end of the Cold War. Our country adopted the nuclear test ban and then developed the ASC program to meet the needs through simulation. However, after a decade of this, we are now starting to transform NNSA's national enterprise of plants and labs to align with our new view of the future. As the Deputy Administrator for Defense Programs Tom D'Agostino recently stated, NNSA is moving towards a leaner, lower cost, more responsive posture that can meet changing requirements coming from the President and Congress. This complex will be called upon to rapidly certify both tested and possibly untested weapons in the future and to be prepared to meet unanticipated threats. These requirements place huge demands on predictive computing.
The second decade of ASC, and beyond, is therefore focused on predictive capability, which is fundamentally about using simulations to help quantify uncertainties that are sufficiently small to be useful in certification. Predictive computing will generate a very large demand for capacity at the capability level.
HPCwire: Why did you pursue BG/L and how is it different than other ASC platforms?
Kusnezov: We really hit a home-run with BG/L. But looking back at how we got here, it was a pretty tough road. There was remarkable skepticism and resistance to this system from both inside the laboratory complex and in the supercomputing community at large. What our labs have been able to do with it now though is beyond compare. Running, debugging and visualizing on more than 100,000 processors is now almost common practice. Today, if you look at the Top500 list, you see how broadly and globally this system is accepted which is a remarkable achievement. The strong support from the Congress and the Department of Energy has been paramount to our success.
Weapons simulations are perhaps the most complex simulations that currently exist. They require a tremendous collection of physics, chemistry, engineering and material science to get a single prediction. Given our need to better understand underlying science and to quantify our level of uncertainty, we recognized about six years ago that new classes of computers were necessary to meet the coming requirement for predictive simulation. We recognized as well that power consumption concerns, the cost of computer room floor space and the slower growth in microprocessor clock frequency improvements would otherwise overwhelm the program's ability to field the necessary computers.
We wanted to find computers that met our requirements at the lowest cost and our advanced architecture program identified the potential of the Blue Gene approach to meet some of our basic science and materials modeling requirements. In addition, LLNL had an excellent decade-long partnership with IBM that had proven itself over the course of several machines.
BG/L's promise of extreme scalability into the petaflops regime, less power, less cost, and a smaller footprint looked very promising. The design uses inexpensive, low power, yet very reliable, embedded processors combined in extremely high numbers (more than 131,000) to achieve huge peak and sustained performance at relatively low cost.
HPCwire: What kinds of ASC applications are being explored with BG/L?
Kusnezov: We are trying to predict the future — the effects of aging on rather complex materials in unusual conditions. In contrast to our other big supercomputers, which focus on the day-to-day upkeep and maintenance of the stockpile, the scientific program on BG/L is exploring how materials age and the effects on their dynamical behavior. As you know, a classical molecular dynamics application investigating Tantalum solidification was the recipient of last year's Gordon Bell Prize for sustaining over 100 teraflops on BG/L. We now have a first-principles molecular dynamics code that just recently sustained 207.3 teraflops, almost twice that of the Gordon Bell Tantalum calculation, and still another first principles code investigating high explosives, super-ionic water, and carbon graphite-to-diamond transitions. That is 207.3 out of 367 teraflops on 131,000 processors for a physically meaningful application code — a remarkable achievement. There are also other materials modeling codes exploring plutonium damage evolution, aging, shock, and phase transitions important to the stockpile.
By the time BG/L was delivered, our program had worked hard for quite some time to put in place a complete research plan with computational and scientific goals going out to the end of its lifecycle. We are squeezing every available cycle out of this computer to get the most return for the program and the nation. There isn't anything we run on BG/L that is not carefully chosen, planned, prepared and executed. When new calculations or ideas develop, we must balance the national security benefit against the work it will displace.
In addition to materials modeling, we've run a variety of large turbulence and instability calculations on BG/L. The goal here is a direct numerical simulation validation of turbulence with applicability to such things as oceanographic and atmospheric inversions, supernovae, and inertial confinement fusion. There's also a newly optimized parallel dislocation dynamics code exploring direct computation of the plastic strength of materials. Many of these codes have successfully used the entire BG/L machine. I should also mention that our ASC alliance centers at University of Chicago, Stanford, University of Illinois and Caltech ran very large turbulence-related calculations last December and January on BG/L.
HPCwire: Is ASC working with others to expand the usefulness of BG/L?
Kusnezov: Absolutely. It is important to move forward together as a country and as federal programs. We are crossing a threshold in computing, where the questions we are now able to ask are much more ambitious and more relevant to our everyday life. Internationally, the NNSA laboratories and the ASC program have blazed the trail, but with the costs of computers coming down many more areas of research will be enabled in the next five to ten years. It is incumbent on us to share our hard-earned experiences with others starting down this path.
Through Lawrence Livermore National Laboratory, we have been participants in the Argonne Blue Gene Consortium since it began. We also have a close working relationship with the Department of Energy's Office of Science for follow-on research and development contracts with IBM for next-generation Blue Gene systems with a far larger applications reach. There is an excellent possibility that most of our codes, including our integrated weapons performance codes, can migrate onto these future Blue Gene systems.
Our work with others has been critical to our success. I mentioned our ASC Academic Alliance Program's use of BG/L. We are continuing our close ties to academia through a new effort called the Predictive Science Academic Alliance Program (PSAAP). This program will focus on unclassified research of interest to NNSA and its national labs — Sandia, Los Alamos and Lawrence Livermore. With the participation of strong academic partners, we can better guarantee our ability to develop the necessary science and engineering applications and uncertainty quantification methodologies that will further establish the viability of predictive science in multi-scale simulations.
On the software side, we continue to participate with the open-source community to develop scalable software for these types of computers. For example, we worked hard to port SLURM (the Simple Linux Utility for Resource Management) and the Lustre parallel file system to the BG/L environment. We've also benefited greatly from the work of others, in particular Argonne's excellent work on MPI for BG/L.
HPCwire: Can you say more about continued inter- and intra-agency cooperation?
Kusnezov: There are many complementary needs in supercomputing and at the federal level we are well integrated. The national needs are quite diverse, from the very pragmatic mission driven needs in NNSA's case where we need very specific tools on fixed schedules, to the open ended research in advanced systems and software critical to ensuring our future capabilities.
We are very pleased to be working with Department of Energy's Office of Science on Blue Gene architecture research and development that promises petaflop and 10 petaflop designs. What we are doing could not be accomplished without close cooperation from the Office of Science and from IBM. This is an ideal government-industry partnership providing advanced computers for the nation that are both preeminent and affordable, while at the same time contributing to science and national security.
The ongoing DARPA HPCS program is another key party providing insight and solutions targeted toward future petaflop class computers. We continue to look at potential petaflop class systems as part of our ongoing advanced architecture efforts and we are benefiting from what is being learned through DARPA's HPCS program and its Phase II partners: IBM, Cray and Sun.
HPCwire: What has surprised you most about your experience with BG/L?
Kusnezov: I was most surprised at how easily it worked and exceeded my already ambitious expectations — in scaling, reliability, power consumption, robustness and usability. Compared to our other systems, it seemed almost “plug and play”. This is not to say that talented people didn't work hard to bring it to this point, but in contrast to other systems we have fielded and with the new territory in computing we were opening, there were no impassable hurdles.
We originally envisioned a scalable molecular dynamics stockpile science research machine. Instead, we got a generally available supercomputer with a far larger applications footprint than we ever imagined, and a much higher than expected number of pleasingly parallel applications that can take advantage of the whole machine. BG/L's predicted “niche” has expanded far past our original, conservative plan. The machine has also been extremely reliable with an MTBF of about six days. This is remarkable, given BG/L's 131,000 processors and five internal networks.
We are immensely pleased with our ability to achieve more than 50 percent efficiency on a real first-principles molecular dynamics code. While BG/L's small memory per node (512MB) has limited its application to our weapons codes, we have found that the fundamental design of BG/L is remarkably amenable to such computing needs with future versions of the architecture.
We also note with interest IBM's success with the Blue Gene system. Blue Gene machines occupy one third of the supercomputing systems in the top 15 of the November 2005 Top500 ranking. These systems are now found in research centers in the United States and around the world, including the Netherlands, Switzerland and Japan.
HPCwire: What do you see as your biggest challenges moving forward?
Kusnezov: Probably our biggest challenge is meeting the requirements coming from transformation in the nuclear weapons complex, especially in developing the means to assess the uncertainty in our computer predictions. Although we are pushing the envelope in multi-scale computational science with remarkable simulations coupled to experimental data, our biggest challenges still lay ahead. Building a predictive multi-scale simulation capability at supercomputing scales in today's world is still pioneering work. Over the next five years, demand for cycles for predictive simulation to meet critical responsiveness needs will increase from tens to hundreds of times what we are capable of providing today.
Another challenge is power consumption as we build larger and larger systems. I don't want to use petawatts for petaflops. Sure this is a bit of an exaggeration, but the power envelope of some of the next generation systems does not make any sense and I would like to work with the vendors to beat this down so that operational costs don't outpace supercomputer investments in this country. It would be a tremendous loss to use the precious funds we have in supercomputing to pay massive power bills. I would really like to see new benchmarks in the Top500 supercomputers that push users to minimize total power usage costs per simulation cycle. I don't worry about peak performance, or the lack of any real efficiency measure for computers, but about the total 'value' to the nation — I like to think of this measure as the product of how much it will cost to buy and maintain the next supercomputer over its lifetime, with the time it takes to solve the problems you built it for in the first place, or 'time-to-solution'. By designing supercomputers to minimize this measure, you can save money and time in delivering the results.
HPCwire: Any last thoughts?
Kusnezov: As our program crosses the “entry level,” 100 teraflops simulation capability goal we set ten years ago, we are experiencing a qualitative change in computing. The leader of a national security effort of a closely allied government, upon seeing work recently done on BG/L, commented, “This changes everything.” By this he meant that simulation at this scale was clearly, and obviously, a tool for discovery that will give America advantage over potential adversaries and is therefore critical to our future national security mission.
While last year's Gordon Bell calculation ran 100 teraflops sustained, and a recent first principles molecular dynamics calculation reached 200 teraflops, speed is not the goal — rather, it is insight. “Getting the job done” is much more important that being number one on a Top500 list. After all, supercomputers are only tools, and the only way to gauge the success of any supercomputer, is to evaluate the results that were obtained during its lifetime and balance that against the total costs. Our program is always thinking several generations of computers ahead with the broader mission challenges in mind, and it is the context of this mission that keeps us focused and driven to succeed. I am not aware of any other program that is so mission driven, scientifically challenging and long-term, that balances national security with leadership science on a schedule. One of the most important jobs for the future of NNSA is to transform the nuclear weapons complex, and simulation will play a vital role in this.