Dynamic Adaptations in ab-initio Nuclear Physics Calculations on Multicore Computer Architectures

Computational resource availability often changes during the course of execution of an application. This is especially true in modern multi-user cluster environments where users can run many high-performance applications simultaneously which share resources such as Processing Elements (PEs), I/O, main memory, network. In such a scenario, it would be greatly advantageous to have applications augmented with adaptive capabilities, particularly during run-time. This involves targeting a computationally intensive part of the application and invoking appropriate adaptations so as to be able to adjust to the dynamically changing system conditions, to prevent drastic performance loss. In this paper, the parallel application MFDn (Many Fermion Dynamics for nuclear structure) used for ab-initio nuclear physics calculations is integrated with a middleware tool for invoking such adaptations. In particular, the multi-threaded Lanczos diagonalization procedure in MFDn is targeted to observe the effect on performance of dynamically changing the number of threads during the iterative process. Performance gains between two to seven times were observed in the presence of competing applications by incorporating these adaptation strategies.