Tuning Master/Worker Applications: A Practical Use Case with MATE

Programming parallel/distributed applications is a difficult task that requires a high degree of knowledge and expertise, especially to achieve the potential performance offered by HPC. Analysis and tuning tools can be helpful for automatically improving applications performance. In particular, dynamic analysis and tuning tools are necessary for applications that vary their behaviour at execution time. MATE is a tool that, employing performance models, can automatically and dynamically tune parallel applications. This work presents how a theoretical performance model has been integrated into MATE for dynamically tune the data distribution and the number of workers of Master/Worker applications. The results show the effectiveness of using performance models for dynamically tuning parallel applications, and the achieved reduction in time when the application modifies its behaviour during its