The aim of this short course is to familiarise users with the hardware and software architecture of the Komondor GPU partition. We will cover the system, node and GPU hardware architecture in detail as well as the modular software environment supporting the development and runtime life cycle of the user projects. Special attention is paid to factors affecting the computational performance, such as the selection of a suitable parallelism model, the cost of data transfer in the memory hierarchy, strategies to reduce overheads. Tools for performance analysis will also be introduced.

It is hoped that the course will help users improve their programs to execute more efficiently and faster, or to be able to execute jobs with increased problem size.