With the increase of demand for computing resources and the struggle to provide the necessary energy, power-aware resource management is becoming a major issue for the High-performance computing (HPC) community. Including reliable energy management to a supercomputer's resource and job management system (RJMS) is not an easy task. The energy consumption of jobs is rarely known in advance and the workload of every machine is unique and different from the others. We argue that the first step towards properly managing power is to deeply understand the power consumption of the workload, which involves predicting the workload power consumption and exploiting it by using smart power-aware scheduling algorithms. Crucial questions are (i) how sophisticated a prediction method needs to be to provide accurate workload power predictions, and (ii) to what point an accurate workload's power prediction translates into efficient power management. In this work, we proposed a method to predict and exploit HPC workloads power consumption with the objective of reducing the supercomputers power consumption, while maintaining the management (scheduling) performance of the RJMS. Our method exploits workload submission logs with power monitoring data, and relies on a mix of lightweight power prediction methods and a classical EASY Backfillling inspired heuristic. Then, we model and solve the power capping scheduling as a greedy knapsack algorithm. This algorithm improves the Quality of Service and avoids starvation while keeping the solution lightweight. We base this study on logs of Marconi 100, a 980-node supercomputer. We show using simulation that a lightweight history-based prediction method can provide accurate enough power prediction to improve the energy management of a large scale supercomputer compared to energy-unaware scheduling algorithms. These improvements have no significant negative impact on performance.

Reducing greenhouse gas (GHG) emissions from Information and Communication Technology (ICT) has become a hot topic since the Paris Agreement. Data centers are one of the most impactful ICT energy consumers since they are built to run 24 hours / 7 days. An emerging discussion is switching their power supply from brown to green energy, using Renewable Energy Sources (RES). However, this change introduces uncertainties linked to production intermittence. This work is part of the Datazero2 project. This project designs a data center powered only by renewable production, adding storage elements to reduce the impact of the intermittence. A clean-by-design data center requires several decisions at different levels of management. To do so, it uses predictions to plan the actions for the next few days. However, it also needs to react to the actual events that can vary from the forecast. This work presents the BEASY heuristic. BEASY mixes power and scheduling decisions in a renewable-only data center, seeking to reduce the number of killed jobs and wasted energy. The results demonstrate that BEASY reduced wasted energy by up to 35.33% in critical cases. Considering the killed jobs, it also kills fewer jobs than the state of art algorithms in all executions.

International audience

Both academics and industry have engaged their efforts in reducing greenhouse gas (GHG) emissions of Information and Communications Technology (ICT). Data centers are one of the most electricity-expensive ICT actors due to their uninterrupted service. Reducing the usage of brown energy or migrating to green energy using renewable sources (RES) is a way to reduce these emissions. However, this migration is not straightforward due to intermittence from these sources. This work is part of Datazero 2 ANR project. This project aims to design a data center powered only by RES production and storage elements. This architecture requires several decisions at different levels of management. This project divides the problem into two groups: offline and online. On the offline side, it uses renewable and workload predictions to prepare an offline plan with a power envelope (power delivered to the servers) and hints on how to manage the storage (batteries and hydrogen). Given the offline plan, the article's contribution is how to deal with real and dynamic power constraints online while keeping the planned storage level at the end. So, this article proposes policies to modify the plan according to the changes in predictions. We evaluate these policies in a homogeneous and heterogeneous data center. The results demonstrate that our policies could approach the storage level and improve Quality of Service (QoS) independently of data center infrastructures.