ALPACA in Florence – first steps towards the Ponte Vecchio”: Porting ALPACA to Intel’s upcoming GPU architecture Ponte Vecchio

Applicant

Prof. Nikolaus A. Adams
Dr. Stefan Adami
Chair of Aerodynamics and Fluid Mechanics (Nanoschock)
Technische Universität München

Project Overview

Over the past seven years, we have developed one of the most advanced in-house simulation frameworks for compressible multiphase flows using multiresolution compression techniques together with the sharp-interface level set method. ALPACA offers a broad collection of state-of-the-art schemes to simulate hyperbolic and parabolic conservation laws. Accurate predictions of, e.g., the Navier-Stokes equations require typically very high temporal and spatial resolution, as hyperbolic problems are governed by multi-scale nonlinear phenomena like shock-propagation and -interaction with complex flow structures. Therefore, adaptive multiresolution compression is employed to minimize computational and memory load on distributed memory architectures via MPI. The challenge for an efficient parallelization is the strong data dependency and complex compute/communication patterns of the fine-grained MR strategy.
The key objective of this KONWIHR proposal is to prepare our in-house simulation framework for next-generation high-performance computing architectures. Whilst much efforts are spent nowadays on CUDA-capable codes, we want to focus in this project on the new Intel GPU architecture “PonteVecchio”. Regarding the programming model, the current available
information is limited. However, ALPACA is well prepared for a MPI+X parallelization strategy, where the local acceleration can be realized with “any” concept. This approach will then be embedded in the hybrid MPI scheme to handle multiple devices of “any” kind. Thanks to the successful and long lasting collaboration with LRZ, we are highly interested in early access and adaption of ALPACA to PonteVecchio, as the upcoming extension of SuperMUC-NG will offer for the first time this architecture. With this transformation from a pure CPU-based framework to a hybrid MPI+GPU(Intel) code, ALPACA will serve for the scientific community and our research as key environment with regards to CFD.