(RP03) HPC with Unstructured Meshes on Novel Architectures
Performance Analysis and Optimization
TimeTuesday, June 26th8:30am - 10am
DescriptionThis poster focuses on many-core parallel solvers for discretized linear and non-linear partial differential equations in the context of cardiovascular applications. Analyses is performed on a set of applications with unstructured discretization based on tetrahedral meshes being considered as part of the co-design process in Mont-Blanc 3. We perform evaluations and analysis on both ARM and Intel based platforms. We also demonstrate how algorithmic changes and different of memory access patterns  reduce the overall energy consumption . Performance tools  identifies fundamental issues that limit performance on the currently available platforms and through predictive studies we identify relevant issues at larger scales. Overall, load imbalance is identified as a very important issue. Operating system noise, interaction between NUMAness and process migration, and memory bandwidth limitations have different impacts depending on the application characteristics. APIs to shift frequency, memory bandwidth or cache capacity between processes allow to compensate small imbalances during runtime. Co-design between the architecture, the operating system and the runtime is the key to achieve high efficiency.