Available courses

In all modern HPC systems, the compute node is where code is executed and "performance is generated." Hence, this is where a deep understanding of the performance issues of any application must start. At first glance, computer architecture appears extremely intricate, making it next to impossible to derive general rules for good performance. However, on closer inspection it turns out that there is a surprisingly small number of guiding principles which govern most of the performance behavior of HPC codes. This online tutorial wants to convey those components of compute node architecture that are most relevant for performance in HPC. We start with the core level and cover code execution via pipelining and out-of-order processing, Single Instruction Multiple Data (SIMD), and Simultaneous Multi-Threading (SMT). Advancing through the memory hierarchy, we look at cache hierarchies, main memory, and cache-coherent non-uniform memory (ccNUMA) architecture. The commonalities and differences between CPUs and GPUs are clearly described. Using simple compute kernels from computational science, we show how architectural features interact with code. We also introduce the Roofline performance model as a simple way to formulate quantitative performance expectations, compare them with observations, and derive possible optimizations. Simple performance tools are introduced that favor insight instead of automation. To make this online event interactive, several online quizzes are interspersed with lectures. Participants can also solve exercise problems using H5P online content and our interactive "Layer Condition Calculator" for stencil codes.

This course, a collaboration of the Erlangen National High Performance Computing Center (NHR@FAU) and the Leibniz Supercomputing Center (LRZ), is targeted at students and scientists with interest in programming modern HPC hardware, specifically the large scale parallel computing systems available in Jülich, Stuttgart and Munich but also smaller clusters.

This course teaches performance engineering approaches on the compute node level. "Performance engineering" as we define it is more than employing tools to identify hotspots and bottlenecks. It is about developing a thorough understanding of the interactions between software and hardware. This process must start at the core, socket, and node level, where the code gets executed that does the actual computational work. Once the architectural requirements of a code are understood and correlated with performance measurements, the potential benefit of optimizations can often be predicted. We introduce a "holistic" node-level performance engineering strategy and apply it to different algorithms from computational science.

Most HPC systems are clusters of shared memory nodes. To use such systems efficiently both memory consumption and communication time has to be optimized. Therefore, hybrid programming may combine the distributed memory parallelization on the node interconnect (e.g., with MPI) with the shared memory parallelization inside of each node (e.g., with OpenMP or MPI-3.0 shared memory). This course analyzes the strengths and weaknesses of several parallel programming models on clusters of SMP nodes. Multi-socket-multi-core systems in highly parallel environments are given special consideration. MPI-3.0 has introduced a new shared memory programming interface, which can be combined with inter-node MPI communication. It can be used for direct neighbor accesses similar to OpenMP or for direct halo copies, and enables new hybrid programming models. These models are compared with various hybrid MPI+OpenMP approaches and pure MPI. Numerous case studies and micro-benchmarks demonstrate the performance-related aspects of hybrid programming.

Hands-on sessions are included on both days. Tools for hybrid programming such as thread/process placement support and performance analysis are presented in a "how-to" section. This course provides scientific training in Computational Science, and in addition, the scientific exchange of the participants among themselves. This course is organized by VSC (Vienna Scientific Cluster) in cooperation with HLRS and RRZE.



Site announcements

There are no discussion topics yet in this forum