Quantum Monte Carlo simulations on ten thousand to a million cores

We are living through a technological revolution, watching the world change as information technology (IT) permeates every area of our lives. Driven by the extraordinary increase in computer power over the past few decades, the IT revolution has advanced to the point that it has become almost impossible to imagine a world without computers. Science has been affected too, perhaps more than most other aspects of society, with computer simulation now playing a central role in almost all research fields. In materials science, in particular, computer simulations based on density functional theory (DFT) have had a huge impact. DFT is a relatively simple but fully quantum mechanical approach capable of providing an accurate description of the microscopic world in many cases, with no inputs other than the identities of the atoms. Applications of DFT to real-life problems span many disciplines, with recent successes including the prediction of novel catalysts, the design of improved batteries, and a better understanding of the temperature and composition of the Earth's core. However, in many room-temperature biological and chemical contexts, the approximations on which real DFT calculations are based are not good enough and the scientific community is calling for more accurate approaches. This project aims to accelerate the development of one such approach, the diffusion quantum Monte Carlo (DMC) method. Unlike DFT, DMC is normally capable of delivering the high accuracy required for most room-temperature biology, chemistry and materials science. Unlike most other accurate methods, DMC is also capable of simulating very large systems, although at a high computational cost. Our main objective is to improve the CASINO DMC code, developed primarily by Richard Needs and his group, to the point that it can be used by physicists, biologists, chemists, earth scientists, and others in much the same way as DFT is used today. CASINO is the world's most widely used DMC code, but like all existing DMC codes it lacks certain features required for real applications in materials science. For example, no DMC code can yet perform quantum molecular dynamics simulations for general systems. One of the aims of our project is to remedy this deficiency.For the past few decades, computers have become faster as processor clock speeds have increased. Today, however, clock speeds are approaching fundamental limits and computers are becoming more powerful only by the inclusion of additional processors. Personal computers often have four or six cores, and some of the supercomputers used for scientific simulations have hundreds of thousands. The bad news is that programming massively-parallel supercomputers is so difficult that the scientific community is being forced to re-think many of its approaches from scratch. The good news is that DMC is one of very few naturally parallel materials-simulation algorithms, and that CASINO already runs efficiently on machines with 10,000 processors. Once we have made the improvements described in this proposal, we are confident that CASINO will run efficiently on the million-core computers of tomorrow. We need to re-work CASINO to harness the power of the future. A few years from now, when petascale computers are becoming more common, the work described in this proposal will make it possible to use DMC to simulate phenomena that today can only be studied at the DFT level, providing the improved accuracy required to take significant steps forward in many areas of science and technology.

In addition to the academic benefits described in the Academic Beneficiaries section, our work will benefit the nation by producing highly-trained personnel. Computational science has become generally accepted as the third pillar of science, complementing and extending theory and experimentation [International Review of Research Using HPC in the UK, EPSRC (2005)] and is widely used in industry as well as academia. Its importance is illustrated by the success of the CASTEP density-functional code, which originated as an academic project at Cambridge University but has been sold commercially since 1995 and accumulated sales of over $20 million, including many sales to industry. Despite the obvious importance of the field, we estimate that the UK produces no more than 100 people per year with a deep and high-level understanding of computational (as opposed to computer) science. The PDRA's employed by this grant will contribute to this trickle of sought-after experts. More generally, now that the era of continuously increasing computer clock speeds is coming to an end, improvements in computational power are beginning to rely almost entirely on increases in parallelism. Over the next 10 years, every enterprise that requires substantial computational power - a list that includes the electronics, pharmaceutical, transport, banking, financial, defence, internet, and energy industries among others - will be forced to come to terms with massively-parallel computing. For the UK economy to remain internationally competitive, it is essential that our research and education systems produce an increasing number of people capable of programming computers with many thousands of cores and using them to solve real-world problems. The PDRA's employed by this grant and the PhD students trained alongside them will be in a position to render an important service to the nation's economy, whether or not they stay in computational science. In the longer run, far beyond the end of this grant, we expect that quantum Monte Carlo methods will also be used in industry, just as density-functional methods are used today. This will not happen any time soon, mainly because quantum Monte Carlo simulations take a thousand or more times longer to run than density-functional simulations, but DFT is not reliably capable of achieving chemical accuracy and is unlikely in our view ever to do so. Since the average thermal energy per molecular degree of freedom at room temperature is only about 25 meV, whereas chemical accuracy is 43 meV/molecule, the long-term value of DFT as a method for investigating room-temperature chemistry and biochemistry may be limited. This leaves the way open for more accurate methods such as quantum Monte Carlo, despite their high computational cost.