Performance Analysis and Consulting

The term High Performance Computing describes it directly: it is about solving technical and scientific problems using systems of high and maximum performance. HPC methodology is the structured and methodical approach to executing these problems on HPC systems and measuring and optimizing the performance achieved.

CLAIX offers an extensive list of tools for software development. The software tools can be made available in different versions via a module environment.

• Compilers for common programming languages translate the program text into executable machine code. An important step in this process is the target architecture-specific optimization. CLAIX offers current compilers from the open source community (GNU, LLVM project) as well as commercial compilers (Intel, NVIDIA), each in several current versions, in order to support the latest functions of programming models such as OpenMP.
• Programming libraries such as MPI and frameworks such as pytorch enable parallel programming across many computing nodes in the cluster or the efficient execution of artificial intelligence models such as neural networks on one or more GPUs. CLAIX offers these tools directly installed or available via containers.
• Debuggers make it possible to search for errors in programs, e.g. by inspecting variable values in the actual program execution. Here too, CLAIX offers the tools of the open source community (GNU) supplemented by commercial products (TotalView, DDT). Of particular importance is the support of parallel programs, which is provided.
• There are also several tools for performance analysis (Score-P with Scalasca and Vampir, Intel Studio) and correctness analysis (MUST, Archer). Performance analysis tools examine the execution of a (parallel) program and show the behavior so that improvement options can be developed. Correctness tools check the execution of a parallel program for possible errors that can occur specifically in parallel programs, such as race conditions and deadlocks.

These and other tools can be found in the module categories "devel" for development, "io" for input-output libraries, "perf" for performance tools and tools for other tools. The available software and possible license conditions are listed and explained in our IT Center Help documentation portal.

CLAIX offers a wide range of ready-to-use software for carrying out technical-scientific simulations, both from open source packages and from commercial providers. The software can be made available via a module environment. For reasons of clarity, we have divided the software packages into the following categories:

• cfd: Computational Fluid Dynamics
• chem: Chemistry
• material: materials science
• math: Mathematics

The available software and possible license conditions are listed and explained in our documentation portal IT Center Help. Our documentation also includes a large number of specific examples of how to use these software packages on CLAIX.

For all users of CLAIX, HPC method support offers assistance in the form of HPC expertise that goes far beyond basic support. The underlying idea of this offer is to provide support in understanding and improving the performance of a program on CLAIX, as well as checking its correctness. These services and the tools used in them are often direct results of research and the challenges faced by users often provide new impetus for research.

The services outlined below are free of charge and are provided within the scope of our available capacities.

Performance assessment

After consultation with the users, we identify a suitable data set for their application, for which a runtime profile is created and various metrics are calculated.

• The runtime profile shows how much time is spent in which parts of the application and how the application behaves on CLAIX. This often reveals a number of effects that limit the execution speed or scaling of a parallel application. Together we consider how these can be eliminated.
• The metrics allow the performance of a code to be compared over time, e.g. across different versions or generations of HPC systems. The template for this is the series of EU Center of Excellence projects POP (Performance Optimization and Productivity), in which RWTH was and is involved. The metrics are described on the POP website.