A model-driven approach for time-energy performance of parallel applications

Summary form only given. Traditionally, time performance is one of the key concerns in running parallel applications. However, the advent of datacenters increases the availability of systems with diverse performance-to-power ratios. These heterogeneous systems consist of hardware including high-performance processors, low-power servers, GPU accelerators, and storage and network resources with different performance. As a consequence, the very large system configuration space introduces new opportunities and complexity for users to run parallel applications efficiently. This keynote discusses a new model-driven approach to determine time-energy efficient configurations for executing applications on heterogeneous systems. By modeling the workload service demands on cores, memory and I/O devices of a node, we obtain the energy-efficient mix of nodes that services a job while maintaining a service time deadline. We show that there is a range of “sweet spots” or Pareto-optimal configurations for executing an application within an energy budget and a given execution time deadline. This keynote is divided into three main parts. First, we review the challenges that users and datacenter providers faced in achieving time and energy efficient parallel application execution. Secondly, we introduce our model-driven approach and the formulation of the mix-and-match execution time model. As an example, we discuss the application of our approach on a heterogeneous system consisting of AMD brawny nodes and ARM wimpy nodes with typical data center workloads including web-hosting, multimedia streaming, financial analysis, real-time speech recognition and TLS/SSL key encryption. In addition, we investigated a number of research questions including is heterogeneity better than homogeneity, are larger mixes of heterogeneous nodes better, among others. Lastly, we highlight new opportunities in modeling time-energy performance such as in cloud computing.