Technical Description: This award from the National Science Foundation to Rutgers University in New Brunswick is for the development of a new computer architecture with multiple levels of interconnected memory, optimized for the simulation of strongly correlated materials from first principles calculations. Strongly correlated materials have the potential to be transformative, two examples: thermoelectric materials, which can generate electricity from waste heat with high efficiency; and, superconductors with potential for higher critical temperatures, fields, and currents, which can revolutionize the electric grid by reducing transmission losses. The use of computational methods to accelerate the pace of discovery of materials with desirable properties is one of the greatest challenges in condensed matter science. Materials with strongly correlated electron systems are particularly difficult to simulat because their physical properties cannot be accurately represented in terms of a system of independent particles moving in an average potential, thus requiring new methodologies and powerful supercomputers for their theoretical description.
Non-Technical Description: This award from the National Science Foundation to Rutgers University in New Brunswick is for the development of a new computer architecture optimized for the dsicovery of new materials from first principles calculations. The new computer will enable collaborations between material synthesis groups and computational physicists. The instrument will drive computer science research, will be used as a resource for teaching computational science and engineering at Rutgers, and will serve as a prototype for a future supercomputer in the national cyber-infrastructure.
We developed a novel computer architecture for big data manipulation, which is based on solid state memory and commercial components assembled to give good performance for electronic structure calculations of correlated electron materials using LDA+DMFT. It is depicted in Fig. 1. It consists of 33 nodes, connected via very fast network interface and switches. Each node has two network cards, one card is designed to communicate with other computers via message passing interface (MPI), while the other is dedicated to passing large amounts data between RAM on motherboard and its extension in the form of solid state memory. We build a large 25 TB Data Staging cloud by adding 800 GB Solid State Drive to each node (NVRAM) and connecting all these drives through a very high speed network. This Data Staging Cloud is a single shared memory space which can be accessed from every node through the parallel file system. Data hungry applications showed a substantial performance gain (200%-900%) on our custom designed instrument relative to a test machine containing a similar number of HDD hardrives. The power of this instrument was demonstrated by LDA+DMFT calculations on V2O3 and Sr2IrO4. In summary, we demonstrated that NVRAM is a promising avenue for improving performance of workflows underlying electronic structure calculations of solid state materials.
