How to Determine the Topology of Hierarchical Tuning Networks for Dynamic Auto-tuning in Large-scale Systems

Abstract

Automatic analysis and tuning is a key strategy that helps to exploit the potential of high performance systems. However, for parallel applications with long running times, dynamic behaviour or highly data dependent performance patterns, it is necessary to make use of the strength of dynamic auto-tuning. An important factor in dynamic auto-tuning on a large scale is the number of additional resources required by the tuning system itself in order to reduce impact on the application performance. A tradeoff must be made between the loss of effectiveness of a tuning system using too few resources and the loss of its efficiency using too many resources. Most automatic analysis or tuning systems do not provide assistance for defining how many additional resources are required. In this work, we address this problem proposing a method focused on calculating the structure of hierarchical tuning networks. The topology will be composed of the minimum number of non-saturated resources. Experimental evaluation performed covers different use cases, each one showing that tuning networks built according to our proposal make efficient use of resources, while providing a high quality analysis and tuning environment.