Pierre Jacquet

Affiliation: Universite du Quebec (ETS Montreal, Canada) and OVHcloud

Talk Title: Monitoring GPUs in Cloud Platforms: Challenges and Opportunities for Orchestration

Abstract: Accelerators are a well-known driver of power consumption growth in cloud infrastructures. Yet, monitoring the individual energy consumption of GPUs remains challenging for providers, due to VFIO behavior and shared resource settings. Such monitoring is nonetheless valuable for providers seeking to study instance-level usage and report on environmental impacts.

In this talk, I focus on IaaS platforms, where GPUs are commonly rented by users to run diverse workloads. Using OVHcloud, a hyperscale cloud provider, as a case study, I detail methods to track the power usage of instances and infer compute usage observed at scale. Building on these findings, I discuss directions for improving GPU allocation through better orchestration of IaaS platforms.

Bio: Dr. Pierre Jacquet is a postdoctoral researcher at Universite du Quebec (ETS Montreal, Canada) and research scientist at OVHcloud, a hyperscale cloud provider. He holds a Ph.D. in Computer Science from the University of Lille, France. His research focuses on large-scale datacenter environments in applied contexts, spanning the full lifecycle of cloud instances from design and orchestration to monitoring. A particular emphasis of his work is reducing the environmental impact of cloud infrastructures.

Homepage: https://jacquetpi.github.io/