Distributed and Parallel Systems Group

OpenCore:

A ManyCore Compiler for Industrial Engineering Stability Analysis

Description

This project aims at solving the computationally-intensive nonlinear structural analysis problem of large and complex light weight structures by using modern parallel multi-core processors enhanced with hardware accelerators such as e.g. GPUs. The project therefore addresses the following two research gaps.

On the one hand, the structural response of lightweight structures is characterised by the occurrence of bifurcation points which determine the load carry capacity. A rigorous treatment of the load carrying behaviour of complex structures consists of three steps: (1) exact localisation of bifurcations in course of the equilibrium path continuation; (2) classification of the singularity and determining the direction(s) of the bifurcated branch(es); and (3) choosing a predictor for initialising the path following procedure in the post-buckling regime. While steps 1 and 2 are well understood from a mathematical point of view, step 3 relies essentially on engineering decisions. However, a reliable and robust assessment of highly loaded, imperfection-sensitive structures is still a challenging problem for the engineer applying commercial finite element programs. The nonlinear analysis of complex lightweight structures like the Front Skirt of the ARIANE 5 launcher with three million degrees of freedom requests a huge computing time. The detection of critical domains within the structural response claims a large part of the computing time using currently available commercial Finite Element Method (FEM) software like ABAQUS or ANSYS.

On the other hand, processor development encountered the three walls for serial performance: the power wall, the memory wall, and the instruction-level parallelism wall. This led to CPUs aggregating multiple processing cores as well as specialized hardware accelerators with very high peak performance gained though their massively parallel architecture. Unfortunately, each of these new devices is offering a different programming interface and execution environment which hinder applications from being ported to different platforms and make the joint use of different tightly-coupled multi-core devices difficult or impossible.

Goal

First, a generic compilation and runtime execution environment for running parallel applications on hybrid multi-core parallel computers will be developed. This environment will include the compilation, optimization, and scheduling of parallel applications written in

OpenCL

on heterogeneous multi-core platforms. A new execution model using a workflow-based intermediate representation of the application which will serve as input to a runtime environment for dynamic optimization and scheduling on such machines. By maximizing data locality and minimizing the load imbalance this environment will accelerate parallel applications.

Second, we will implement enhanced algorithms that use the potential of many-core computing to improve the performance and accuracy of structural analysis. Today, it is widely accepted that future developments of aerospace structures will be more and more realised by computer simulations. The model size of the complete system and of substructures will increase tremendously in order to fulfil the mass targets accompanied by elevated margins. In this context, branching analysis represents the core of a new software family developed by INTALES and four departments of the University of Innsbruck: Engineering Mathematics, Applied Mechanics, Mathematics, and Computer Science. With the parallelisation of the bifurcation and limit point analysis, INTALES will bypass the commercially available analysis software mainly built in FORTRAN with concurrent improvement of the analysis quality. The current experience from the finite element analysis of large launcher structures with qualified software is that up to 50% of the computational time is consumed for the branching analysis. This time shall be minimised with improved quality of results

Workflow of the Patent categorization task.

Illustration of the execution and runtime environment to be developed.

Partners

 

Staff

Radu Prodan

Ivan Grasso

Klaus Kofler

Clemens Domanig

Thomas Fahringer

Publications

Sponsor

This Project is sponsored by the  Österreichische Forschungsförderungsgesellschaft mbH (FFG)