SONAR: Automated Communication Characterization for HPC Applications

Abstract

Future computing systems will need to operate within hard power and energy constraints, this is particularly true for Exascale-class systems. These constraints are hard for technical, economical and ecological reasons, thus, such systems have to operate within given power and energy budgets. Therefore, we anticipate the need for modeling tools that help to predict power and energy consumption. In particular, such modeling tools would allow for detailed explorations of various alternatives when designing systems. While processing and memory already receives a large amount of interest from the research community, power modeling of scalable interconnection networks is rather neglected. However, analyses show that the network contributes about 20 % to the overall power consumption of HPC systems. Considering the increasing energy efficiency of other components, this fraction is likely to increase. While models for processing and memory typically rely on performance counters to model power and energy, we observe that the distributed nature of networks leads to significantly more complex metrics. Selecting the right set of abstract metrics, which will be used as input for such a prediction, is crucial for prediction performance.

In this work we introduce our tool called Simple Offline Network Analyzer (SONAR) to derive complex metrics from communication traces of HPC applications. We explain the motivation behind choosing this concept, the implementation, and the ability of the tool to easily support the integration of new metrics. We also show exemplary explorations using an initial set of metrics for a representative range of HPC applications, including contemporary as well as emerging Exascale workloads. In particular, we use SONAR to characterize the communication of applications in terms of verbosity and network utilization, as we believe both to be important metrics for power prediction.