This project is an experimental and analytic investigation of a previously untested type of network interface as the interconnect of a cluster of workstations for the calculation of atomic structure data. As an example, recent calculations for the prediction of the electron affinity of boron yielded an answer that spurred new experimental measurements of a boron atom. However, such accuracy for larger atoms may present a considerable computational challenge. The hardware platform being studied is a cluster of workstations interconnected with a network interface that contains user-level accessible memory. Each node has memory in the network interface that can be used to cache shared data, and that may be read or written by any node.
Primary efforts are to (1) propose modifications to the atomic structure calculation that use network interface memory as a respository for selected shared data, (2) evaluate via analysis and experimentation the effectiveness of using shared buffers and zero-copy messaging for atomic structure calculations, and evaluate tradeoffs such as memory partitioning techniques and data and thread placement, and (3) propose generalizations of these experimental results for use in other scientific code.
It is expected that investigations will lead to a new proposal for an implementation of large atomic structure calculations with better performance.