A call to arms for task parallelism in multi‐scale materials modeling

Abstract

SUMMARYSimulations based on multi-scale material models enabled by adaptive sampling have demonstrated speedup factors exceeding an order of magnitude. The use of these methods in parallel computing is hampered by dynamic load imbalance, with load imbalance measurably reducing the achieved speedup. Here we discuss these issues in the context of task parallelism, showing results achieved to date and discussing possibilities for further improvement. In some cases, the task parallelism methods employed to date are able to restore much of the potential wall-clock speedup. The specific application highlighted here focuses on the connection between microstructure and material performance using a polycrystal plasticity-based multi-scale method. However, the parallel load balancing issues are germane to a broad class of multi-scale problems.

The use of discrete harmonics in direct multi-scale embedding of polycrystal plasticity

“…For now we employ Eqs. (23) and (24) to demonstrate the approach with a comparatively simple formulation. While the implicit solve of Eq.…”

“…These techniques have some commonalities with in situ adaptive tabulation and similar schemes developed for modeling chemical kinetics [20][21][22]. Previous versions of the embedded polycrystal plasticity with adaptive sampling did not include texture evolution [17,23]. The primary thrust in the present work is to allow for texture evolution, including the effects of both crystallographic slip and deformation twinning, in a tractable multi-scale approach.…”

“…Recall from Section 2 that the single harmonic coefficient h 0 for d pf = 0 is constant and does not contribute to n h . These simulations are conducted using MPMD parallelism to decompose the workload, as in [18,23]. As noted in Section 3.3, this enables more rapid evaluation of f Y (Eq.…”

“…The fine-scale model provides plastic deformation response and updated material state under these conditions. While the input dimension of the response space may be large, the portion of the response space probed in a given calculation tends to fall on a low-dimensional manifold [21,23]. The points on this manifold are not known ahead of time, motivating the use of an in situ adaptive sampling method that computes responses as needed.…”

“…(24) as described below. In Section 4 we touch again on the decomposition of the workload onto computational resources, but we refer the interested reader to [23] for more details related to the use of computational resources. As described below, most of the calls to the adaptive sampling module result in an approximation of the fine-scale response based on previous evaluation data and do not require a new fine-scale model evaluation.…”