Best practice for efficient and scalable application performance

Nowadays, the whole HPC community is looking forward to the exascale era, with computer and system architects investing their effort to design the first exascale supercomputer. The race winner is as uncertain as ever, looking at the diversity in architectures, paradigms, and vendors at the top of the top500 list.On the other hand, application developers and scientists from different fields relying on HPC struggle to prepare their codes to run efficiently, even on current leadership computer systems. Some of the challenges they face are well known (e.g., parallelization inefficiencies caused by synchronizations, load imbalance, or communication patterns). Other constraints/inefficiencies arise from new computing paradigms (e.g., how to manage input/output when dealing with big data) and the heterogeneity of computing resources (e.g., when/how to efficiently exploit an accelerator).After more than 200 assessments including flagship codes of Centres of Excellence (CoEs) addressing a variety of scientific domains, the Performance Optimisation and Productivity (POP) CoE has a unique view of the most common pitfalls and sources of inefficiency that prevent codes from running efficiently and achieving scalability. We present a set of common patterns found in parallel codes that lead to performance loss. By detecting and isolating these patterns, we make them independent of the code and the scientific algorithms behind them. Developers from any domain can recognize these patterns in their codes and be alerted to their performance issues.We also share recommendations on how to address these patterns. The best practices proposed are not only oriented to improve the codes' performance but also to maintain developer productivity and portability of the codes. These will be crucial in the exascale race to cope with the increasing complexity of applications and computer systems, deal with the variety of architectures and paradigms, and ultimately run efficiently in an exascale supercomputer.Both the patterns and the recommendations are presented in a generic and understandable manner to reach a broad audience from any scientific domain targeting HPC systems