P1 - D3.3 Report on Exa-enabling enhancements and benchmarks

Type

Public material

Description

The document summarizes the main progress and achievements on the development of Exascale enabling technologies on the EXCELLERAT core codes during the life of the project, but with focus on the last year. Previous reports, D3.1 and D3.2, contain detailed information about the activities conducted in year 1 and 2 respectively. The developments have been driven by the definition of individual code development roadmaps in collaboration with WP2 and WP4 to demonstrate Exascale simulations for the use-cases.

From this roadmap, several requirements were identified (see D2.1, D2.2 and D2.4 about ”Reference_Applications_Roadmap and Challenges”) and a summary of the activities conducted to address these requirements is presented here. Two fundamental activities are associated with these developments: i) Task 3.1 focused on node-level performance and ii) Task 3.2 on system-level performance engineering. Note that main changes in the evolution of HPC systems are occurring at node level. This is a major reason to have a specific task focused on this topic.

In this final year, the activities carried out by the partners have been focused on the development of the application demonstrators of the use-cases, mainly by the use of GPUs, emerging technologies and the use of Adaptive Mesh Refinement (AMR). At node level (Task 3.1), analysis and optimization of the linear solver Spliss on GPU architectures was conducted by DLR for CODA, and DLB in Alya. At the system level (Task 3.2), the focus has been on strong scaling analyses and on the optimization of the communication kernels. Regarding the advanced meshing techniques (Task 3.3), most of the work has been performed on the different codes from the project. It includes m-AIA with a mesh adaptive level-set method combined with dynamic load balancing, a parallel AMR strategy based on Space Filling Curves for mesh partitioning for Alya, TREEPART with domain decomposition library with dynamic load balancing for AVBP and AMR for Nek5000 using different graph partitioners ParMETIS and PARRSB. The advances in HPC algorithms and computational methodologies presented here are part of the expertise of the EXCELLERAT consortium and compile a set of services that EXCELLERAT is delivering to the engineering community.

Web-URL

Please find the Deliverable here. 

This deliverable was developed within the EXCELLERAT P1 Project phase.