Labs Paper on Energy Aware Architectures Earns Award

[Labs_Editorial]

Of all the research challenges Hewlett Packard Labs takes on, one of the most pressing is trying to make data centers more sustainable. The planet, after all, is facing a climate-change crisis, and data centers are contributing to the problem. They account for 0.6% of the world’s greenhouse gas emissions and 1% of its world’s electricity, and they use inordinately large volumes of water.

Reducing data centers’ carbon footprint won’t get society to Net Zero targets all by itself. Cutting emissions from cars, planes, factories and utility grids would move the needle more forcefully, but data center improvements can make a difference. So, Labs is pursuing initiatives to make data centers more energy efficient by making their architectures more “energy aware.”

“Energy aware architectures enable multiple designs and implementations to account for energy awareness in a coherent manner, all deriving from the same architectural principles,” said HPE Fellow Dejan Milojicic. “This enables consistency across our products and improves interoperability. But the most important aspect is that they collectively optimize energy usage.”

Labs’ work on energy aware architectures is starting to turn some heads. A paper Labs researchers and university collaborators published at the International Green and Sustainable Computing (IGSC) Conference in October was recognized as the Best Paper at the conference.

The paper, “SHIELD: Sustainable Hybrid Evolutionary Learning Framework for Carbon, Wastewater, and Energy-Aware Data Center Management,” was a joint development project between researchers at Colorado State University (CSU) and at Labs. It was co-written by Sirui Qi and Professor Sudeep Pasricha from CSU and by Milojicic and Cullen Bash from Labs.

“Like most universities we collaborate with, CSU was responsible for out-of-box thinking and coming up with innovation unaffected by corporate constraints,” Milojicic said. “We then apply these constraints to make sure we have a practical and workable solution. In the process both sides win and most of all our customers benefit from this collaboration.”

Managing Workload Distribution

The initiative proposes a new way to co-optimize carbon emissions, water footprint, and energy costs of globally distributed data centers (GDDCs). CSU and Labs co-developed a hybrid workload management framework called SHIELD that uses machine learning to intelligently manage workload distribution across GDDCs and data center operation.

The proliferation of globally distributed data centers to handle exponentially growing data volumes has put a higher priority on reducing the energy costs and environmental (water, carbon) overheads of data centers. One way to do this is to create a system that assigns workloads to data centers where energy is cheaper and cleaner. These two perspectives are usually in conflict, so it puts the burden on cloud service providers to manage workloads judiciously.

Labs’ SHIELD project combines machine learning and evolutionary algorithmic techniques to co-optimize the three sustainable objectives for GDDCs: minimizing carbon emissions, water footprint, and energy costs.

Results show that SHIELD outperforms other state-of-the-art data center management frameworks in several key areas, including convergence speed, scalability and solution quality. Specifically, carbon footprint was reduced by up to 3.7×, water footprint by up to 1.8×, and energy costs by up to 1.3×.

“Sustainability has multiple dimensions, and only multi-objective solutions will address all dimensions compliant with customer needs and with infrastructure capabilities,” Bash said. “No data center is an island, and our work leverages time differences to follow the availability of clean energy. We get clean energy more efficiently by executing some of the workloads where energy is less carbon-intense and cheaper at a given time.”

Other Labs research projects aimed at data center sustainability include digital twins, an energy-aware scheduler and architectures that enhance energy-awareness using silicon photonics and accelerators. Researchers also are studying prospects for energy-aware architectures that scale from edge to core to cloud.

In the next phase, Labs will focus on profiling serverless Functions as a Service (FaaS) which have some unique characteristics, such as stateless that enables easier offloading.

“We are exploring energy footprints of different FaaS, especially for accelerators,” Milojicic said. “Serverless and accelerators represent a good match due to the fine granularity of the two solutions.”