Modeling for Performance Engineering: the Case of Graph Processing

Graphs are universal abstractions, often used to represent concepts, objects, or individuals and the relations between them. When graphs are used to solve real-life problems – like the analysis of emerging inter-molecular interactions, the efficiency and feasibility of logistics networks, the reliability of support networks, or the safety of social networks – they quickly become massive in scale and/or complexity, and thus require massive computational resources to be processed. In turn, such large-scale processing of graph workloads raises efficiency and sustainability concerns: the irregular, data-intensive nature of graph processing algorithms often leads to significant computing waste. To alleviate these concerns, we propose a co-design methodology to enable the selection of efficient graph-processing algorithms _and_ their effective deployment on suitable infrastructure. Our approach relies on design-space exploration, driven by efficient search methods and compositional performance models. Our models take both workload and data features into account, and their composition enables complex workloads analysis. We showcase the impact of such modeling to reduce the workload’s energy consumption, while increasing resource utilization with limited performance impact.