Parallel Matrix-Free Methods for Simulating Spin-1/2 Hamiltonians
Julia Wei, Harvard University
Spin models are used to study phenomena in quantum physics, such as magnetism and phase transitions. Dynamite is a python package with a C backend that simulates quantum spin models. In particular, it can compute the energy eigenstates of a spin-1/2 Hamiltonian and the time-evolution of quantum states. In the backend, dynamite calls Krylov-subspace methods in external libraries (SLEPc and PETSc), where the workhorse is matrix-vector multiplication. We have implemented two matrix-free features, both of which use MPI parallelization: 1) a matrix-vector product for Hamiltonians that conserve total magnetization, and 2) a CUDA kernel that can run on multiple GPUs. We show performance results for these features, for different spin models and system sizes.
Abstract Author(s): Julia Wei, Greg Kahanamoku-Meyer, Norman Yao