Porting Code to the GPU promises large speedups, but there can be pitfalls on the way to realizing its potential. This course will introduce NVIDIA's profiler as a tool to spot common performance bugs that arise when porting code to the GPU.
A practical demonstration introduces the basics of GPU performance analysis. NVIDIA's profiling tool Nsight Systems ist used to analyze GPU utilization and to spot performance anomalies. The complementary tool Nsight Compute is then used to gain more insight about the performance of individual GPU kernels. The performance analysis will be guided by simple, resource-based performance models which will enable the developer to develop a concept of how far the performance is from the "target."
Attendees will be able to follow along the demos and conduct their own experiments on the NHR@FAU GPU cluster.
Requirements
- basic understanding of GPU programming with CUDA, OpenMP, or OpenACC
- compile applications by invoking the compiler on the command line
- editing files through ssh connection via vim, emacs, nano, ...
Participation
Participation is free of charge for attendees from German universities, academic computing centers, and research institutions. Please only register for the course if you are really going to attend. No-shows will be blacklisted and excluded from future NHR@FAU events.
Certificates
All actual course attendees will receive a course certificate.
Instructor
Dominik Ernst (NHR@FAU)
This course is organized by Erlangen National High Performance Computing Center (NHR@FAU)
