No Disk Left Behind: Lifetime-Aware Oversubscription of Thin Provisioned Storage in Cloud IaaS platforms, P. Jacquet, C. Coti, M. Dias de Assunção International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2026), 2026 [Code]
Cloud resource oversubscription is widely used to improve infrastructure utilization, particularly for CPU. In contrast, storage resources remain largely statically provisioned despite significant underutilization: our analysis of multiple OVHcloud platforms shows that virtual machines use on average only 42.9% of their local disk capacity. This paper explores disk oversubscription in thin-provisioned environments, where storage resources are persistent and difficult to predict in the black-box context of IaaS platforms. We introduce Disk Release Prediction (DRP), a dynamic oversubscription method that anticipates disk capacity released by instance churn. Using large-scale traces from OVHcloud, we show that DRP can reduce required disk capacity by up to 2x while maintaining a low rate of corrective migrations, lowering storage-related scope 3 emissions by approximately 44.5% in newer cloud regions.
@inproceedings{scdisk,
title = {No Disk Left Behind: Lifetime-Aware Oversubscription of Thin Provisioned Storage in Cloud IaaS platforms},
author = {Pierre Jacquet and Camille Coti and Marcos Dias de Assunção},
booktitle = {International Conference for High Performance Computing, Networking, Storage and Analysis},
year = {2026},
url = {https://zenodo.org/records/21262313},
}
Parallel Hybrid Quantum-Classical Computing: Programming Model and Platform Simulation, L. Brochu, F. Bureau, P. Gasnier, P. Jacquet, C. Coti IEEE International Conference on e-Science (eScience), 2026
Hybrid quantum-classical applications require execution environments able to coordinate conventional computing resources and quantum processing units (QPUs). As quantum platforms remain scarce and difficult to access for systematic scheduling studies, simulation and runtime support are needed to evaluate programming models, performance models, and scheduling policies before deployment on physical hardware. This paper presents an open-source framework for studying hybrid quantum-classical execution: we extend SimGrid with QPU resources able to execute quantum circuits through an emulator while injecting platform-specific execution-time estimates, and StarPU with a task-based programming model in which quantum tasks can be scheduled alongside classical tasks. Empirical measurements on the IBM ibm_quebec quantum computer, integrated into the simulator, show that circuit execution time is mainly driven by the number of shots, while circuit depth and number of active qubits have a weaker observable influence under the tested conditions.
@inproceedings{esquantum,
title = {Parallel Hybrid Quantum-Classical Computing: Programming Model and Platform Simulation},
author = {Linden Brochu and Francis Bureau and Pacôme Gasnier and Pierre Jacquet and Camille Coti},
booktitle = {IEEE International Conference on e-Science},
year = {2026},
}
CINERGY: Deterministic Power Monitoring for Carbon Accounting in the Cloud, P. Jacquet, C. Coti, M. Dias de Assunção, R. Rouvoy IEEE Transactions on Cloud Computing (TCC), 2026 [Paper][Code]
Cloud providers now commonly offer tools to monitor the environmental impact of hosted services, a significant share of which arises from the power consumption of servers. To allocate server-related emissions, these methodologies typically rely on resource-quantity-based attribution and do not account for the effective usage of virtual resources. We demonstrate that resource usage can drastically affect carbon estimations, and propose CINERGY, a framework to standardize how cloud providers integrate the power consumption of virtual resources into their carbon calculators. It defines a deterministic baseline for the power consumption of a provisioned VM (same usage, same estimate) while still accounting for consolidation gains enabled by virtualization. Constructed and evaluated empirically, our approach achieves high precision, with a mean absolute error of 6.6% across various hardware. Using metrics collected from cloud providers, we reveal the impact of instance type on energy efficiency (up to 85% differences for the same VM size) and show that quantity-based accounting can underestimate CO2 emissions by up to a factor of three.
@article{cinergytcc,
title = {CINERGY: Deterministic Power Monitoring for Carbon Accounting in the Cloud},
author = {Pierre Jacquet and Camille Coti and Marcos Dias de Assunção and Romain Rouvoy},
journal = {IEEE Transactions on Cloud Computing},
year = {2026},
pages = {822-833},
doi = {10.1109/TCC.2026.3674370},
url = {https://hal.science/hal-05551692},
}
Larger Cloud Servers, Fewer Hosts? On the Evolution of VM Sizes in IaaS Platforms, P. Jacquet, C. Coti, M. Dias de Assunção IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2026
In contrast to hardware, the evolution of the virtual resources through which most practitioners interact with computing infrastructure remains largely unexplored. We study how Virtual Machine (VM) configurations have evolved across major cloud providers over the past two decades, using historical catalogs and large-scale usage traces. Our analysis shows that VM evolution diverges from underlying hardware trends: although provider catalogs have expanded toward larger configurations, client demand remains concentrated on the smallest options. At the same time, larger options are increasingly selected, and gradual increases in minimal allocations have raised the footprint of active VMs, which doubled between 2019 and 2025. Consequently, the number of instances consolidated per host has remained stable over time. By progressively inflating default VM footprints, cloud offerings weaken incentives for software efficiency and limit the potential gains from consolidating workloads on more capable servers.
@inproceedings{cloudevolutionccgrid26,
title = {Larger Cloud Servers, Fewer Hosts? On the Evolution of VM Sizes in IaaS Platforms},
author = {Pierre Jacquet and Camille Coti and Marcos Dias de Assunção},
booktitle = {IEEE International Symposium on Cluster, Cloud and Internet Computing},
year = {2026},
pages = {287-296},
doi = {10.1109/CCGrid68966.2026.00037},
}
Untangling GPU Power Consumption: Job-Level Inference in Cloud Shared Settings, P. Jacquet, M. Agusti, E. Caron, C. Coti, M. Dias de Assunção, L. Lefèvre, A. Orgerie European Conference on Computer Systems (EuroSys), 2026 [Paper][Code]
As the demand for AI-driven workloads increases, the energy consumption of Graphics Processing Units (GPUs) has come under intense scrutiny, particularly in hyperscale data centers where large numbers of accelerators are centralized and leased to diverse clients. However, GPU power monitoring in cloud environments presents several challenges, as the monitoring capabilities of physical devices may be limited or even absent depending on the product offered. We propose methods to estimate the energy consumption of jobs running on GPU devices in cloud environments, spanning shared and managed offerings (ML-as-a-Service) to less managed products (Infrastructure-as-a-Service). Our models demonstrate the benefits of sharing GPUs for small AI workloads, as well as the current sub-optimal utilization of GPUs in cloud hyperscalers, based on insights from an IaaS GPU cluster.
@inproceedings{eurosys26,
title = {Untangling GPU Power Consumption: Job-Level Inference in Cloud Shared Settings},
author = {Pierre Jacquet and Maxime Agusti and Eddy Caron and Camille Coti and Marcos Dias de Assunção and Laurent Lefèvre and Anne-Cécile Orgerie},
booktitle = {European Conference on Computer Systems},
year = {2026},
pages = {624-640},
doi = {10.1145/3767295.3769333},
url = {https://hal.science/hal-05291033},
}
The Case for Time-Shared Computing Resources, P. Jacquet and Adrien Luxey-Bitri Computing Within Limits (LIMITS), 2025 [Paper]
Although the sector appears virtual, Information and Communication Technologies remain inherently constrained by their materiality and cannot rely on an infinite pool of resources. Contrary to common assumptions, we show that tenants typically do not share computing resources, even in environments commonly perceived as mutualized such as cloud platforms: time-sharing has been progressively phased out for reasons of performance, security, predictability, and the decreasing cost of computing resources. This paper advocates for managing fewer physical resources by improving resource sharing between tenants, a paradigm shift moving beyond traditional time-sharing at the hardware level to a higher abstraction. We review the current state of the art, identify challenges and opportunities, and outline key research directions for Time-Shared Computing.
@inproceedings{sharinglimits,
title = {The Case for Time-Shared Computing Resources},
author = {Pierre Jacquet and Adrien Luxey-Bitri},
booktitle = {Computing Within Limits},
year = {2025},
pages = {1-8},
url = {https://hal.science/hal-05130125},
}
CINERGY: Reasoning over the Worst Case Power Consumption of Cloud Virtual Machines, P. Jacquet, C. Coti, M. Dias de Assunção, R. Rouvoy IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2025 [Short Paper][Code]
While the power consumption of physical servers can be physically monitored, the adoption of virtual environments such as cloud computing renders physical measurements impractical in operational contexts. State-of-the-art approaches mostly assign shares of server power consumption to hosted processes using system metrics, but this bottom-up approach is highly sensitive in multi-tenant environments: the same VM activity may lead to different power traces depending on co-hosted VMs. As cloud customers only control their provisioned virtual resources, we propose CINERGY, a method to model the power consumption of virtual appliances with context-agnostic tracking of their environmental impact. Designed to be more predictable than state-of-the-art power models while still exposing the gains from consolidation, CINERGY is deterministic and accurate, with an average error of 6.6% evaluated against ground-truth measurements.
@inproceedings{jacquet:hal-04981001,
title = {CINERGY: Reasoning over the Worst Case Power Consumption of Cloud Virtual Machines},
author = {Pierre Jacquet and Camille Coti and Marcos Dias de Assunção and Romain Rouvoy},
booktitle = {IEEE International Symposium on Cluster, Cloud and Internet Computing},
year = {2025},
pages = {586-589},
doi = {10.1109/CCGRID64434.2025.00024},
url = {https://hal.science/hal-04981001},
}
SlackVM: Packing Virtual Machines in Oversubscribed Cloud Infrastructure, P. Jacquet, T. Ledoux, R. Rouvoy IEEE International Conference on Cluster Computing (CLUSTER), 2024 [Paper][Code]
Cloud providers expose a large catalog of Virtual Machine (VM) offers, some premium with guaranteed dedicated resources and others hosted in oversubscribed environments. Managing multiple oversubscribed VM levels often leads providers to provision isolated clusters of Physical Machines (PMs) for each category of offers. We introduce SLACKVM, a shared architecture in which VMs from various oversubscription levels coexist on the same cluster of PMs. We demonstrate that oversubscription levels can be complementary, meaning they do not saturate the same resource components: by coupling multiple levels, Cloud providers can better consolidate VM offers onto PMs and reduce the size of their clusters by up to 9.6%. These resource savings yield both operational cost reduction and a reduced ecological footprint, with a limited impact on the Quality of Service.
@inproceedings{jacquetslack,
title = {SlackVM: Packing Virtual Machines in Oversubscribed Cloud Infrastructure},
author = {Pierre Jacquet and Thomas Ledoux and Romain Rouvoy},
booktitle = {IEEE International Conference on Cluster Computing},
year = {2024},
pages = {190-201},
doi = {10.1109/CLUSTER59578.2024.00024},
url = {https://hal.science/hal-04636648},
}
SweetspotVM: Oversubscribing CPU without Sacrificing VM Performance, P. Jacquet, T. Ledoux, R. Rouvoy IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2024 [Paper][Code]
VMs schedulers mostly implement all-or-nothing oversubscription strategies, wherein all VM resources are either oversubscribed or remain unaltered. An empirical study of individual virtual CPU (vCPU) usage in the OVHcloud production environment demonstrates that vCPUs are not uniformly utilized, making this holistic approach questionable. We introduce SWEETSPOTVM, where oversubscription ratios are applied at the granularity of individual vCPUs instead of whole VMs, enabling a more flexible oversubscription management strategy by pinning oversubscription ratios per vCPU within VMs. Experiments on a physical platform show the possibility of dividing the cost of hosting VMs by 3 while maintaining VM performance at the level of non-oversubscribed platforms, opening new avenues to boost consolidation with positive impacts on the environmental footprint of cloud computing.
@inproceedings{jacquet:hal-04454043,
title = {SweetspotVM: Oversubscribing CPU without Sacrificing VM Performance},
author = {Pierre Jacquet and Thomas Ledoux and Romain Rouvoy},
booktitle = {IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing},
year = {2024},
pages = {148-157},
doi = {10.1109/CCGrid59990.2024.00026},
url = {https://hal.science/hal-04454043},
}
SCROOGEVM: Boosting Cloud Resource Utilization with Dynamic Oversubscription, P. Jacquet, T. Ledoux, R. Rouvoy IEEE Transactions on Sustainable Computing (TSUSC), 2024 [Paper][Code]
Despite continuous improvements, cloud physical resources remain underused, severely impacting the efficiency of these infrastructures. IaaS providers usually compensate for oversized Virtual Machines (VMs) by offering more virtual resources than are physically available on a host, but statically-defined oversubscription ratios may hinder performance under resource contention. This article studies how a greedy increase of the oversubscription ratio per Physical Machine and resource type can preserve performance goals. Our contribution, SCROOGEVM, leverages the detection of PM stability to carefully increase the associated oversubscription ratios. Based on metrics shared by public cloud providers, our comparative analysis shows that SCROOGEVM outperforms state-of-the-art oversubscription computations by leveraging long-lasting VMs, while avoiding live migration penalties and performance impacts for stakeholders.
@article{jacquet:hal-04466538,
title = {SCROOGEVM: Boosting Cloud Resource Utilization with Dynamic Oversubscription},
author = {Pierre Jacquet and Thomas Ledoux and Romain Rouvoy},
journal = {IEEE Transactions on Sustainable Computing},
year = {2024},
pages = {754-765},
doi = {10.1109/TSUSC.2024.3369333},
url = {https://hal.science/hal-04466538},
}
CLOUDFACTORY: An Open Toolkit to Generate Production-like Workloads for Cloud Infrastructures, P. Jacquet, T. Ledoux, R. Rouvoy IEEE International Conference on Cloud Engineering (IC2E), 2023 [Paper][Code]
Simulators and benchmarks are broadly adopted to study cloud infrastructures, but their reported results may suffer from a representativeness problem, as they vary depending on input workloads whose characteristics differ greatly between cloud providers and public/private deployments. Addressing this representativeness requires providers to share production-scale datasets, which might be considered sensitive. This paper introduces CLOUDFACTORY, an IaaS workload generator composed of a library allowing providers to share IaaS statistics instead of raw datasets, and of a generator producing realistic VM workloads matching these statistics. As an example, we analyze the scheduling evolution for different IaaS workload intensities of two cloud providers (Microsoft Azure and Chameleon), and report OVHcloud statistics computed with CLOUDFACTORY compared to other providers.
@inproceedings{jacquet2023cloudfactory,
title = {CLOUDFACTORY: An Open Toolkit to Generate Production-like Workloads for Cloud Infrastructures},
author = {Pierre Jacquet and Thomas Ledoux and Romain Rouvoy},
booktitle = {IEEE International Conference on Cloud Engineering},
year = {2023},
pages = {81-91},
doi = {10.1109/IC2E59103.2023.00017},
url = {https://hal.science/hal-04168667},
}
Press
La chasse au gaspillage dans le cloud et les data centers, P. Jacquet, T. Ledoux, R. Rouvoy The Conversation, 2023 [Press article][Ouest-France]
@misc{theconversation,
title = {La chasse au gaspillage dans le cloud et les data centers},
author = {Pierre Jacquet and Thomas Ledoux and Romain Rouvoy},
journal = {Ouest-France},
year = {2023},
url = {https://theconversation.com/la-chasse-au-gaspillage-dans-le-cloud-et-les-data-centers-196669},
}
Agir face aux enjeux environnementaux, la thèse comme levier, P. Jacquet Docteurs SPI, 2025 [Article]
@misc{redocspi,
title = {Agir face aux enjeux environnementaux, la thèse comme levier},
author = {Pierre Jacquet},
journal = {Docteurs SPI},
year = {2025},
url = {https://www.docteurs-spi.org/post/agir-face-aux-enjeux-environnementaux-la-th%C3%A8se-comme-levier},
}
Theses
Enhancing IaaS Consolidation with Resource Oversubscription, P. Jacquet Ph.D. Thesis, 2024 [Thesis]
By running workloads in shared data centers, the digital realm has improved its energy efficiency, as visible through indicators such as Power Usage Efficiency (PUE). However, infrastructure is not the sole point of optimization: from both an energy and material standpoint, a server loaded at 100% is preferable to three servers loaded at 30%, making usage rates a considerable potential gain for research. This thesis examines these usage rates along four complementary contributions: (1) the creation of realistic controlled experiments in an IaaS context, through a dedicated workload generation tool; (2) the calculation of individual server oversubscription ratios by considering individual server stability, fine-tuning ratios without additional violations; (3) the introduction of a new oversubscription paradigm based on the observation that VM vCPUs are not uniformly used, exposing cores of different power levels to improve overall performance; (4) the complementarity of oversubscription techniques, since premium and oversubscribed VMs saturate their hosts' resources differently, enabling synergies that reduce the number of servers by up to 9.6%.
@phdthesis{phdthesis,
title = {Enhancing IaaS Consolidation with Resource Oversubscription},
author = {Pierre Jacquet},
school = {Université de Lille},
year = {2024},
url = {https://theses.fr/2024ULILB015},
}
Mainframe footprint and consumption, an evaluation based on sustainable IT approach, P. Jacquet Master Thesis, 2021 [Thesis]
Mainframes remain a key infrastructure element for the banking sector: according to IBM, 92 of the top 100 banks use mainframe computers and 90% of all credit card transactions are performed on mainframes. While sustainable IT has been extensively studied on common platforms such as x86 computers, there are, to the best of our knowledge, no dedicated studies considering the mainframe ecosystem through a consumption prism. We first review consumption evaluation methods (hardware, hybrid, and software approaches) and how to adapt them to mainframe specificities. We then determine the consumption and carbon footprint of the Groupe BPCE mainframe servers, including external components such as tape libraries and flash storage arrays, and compare this order of magnitude against an equivalent x86 architecture. Finally, we study how mainframe specificities, including its architecture and retro-compatibility capacity, could support a sustainable strategy, and outline directions for such a strategy based on identified internal and external factors.
@mastersthesis{masterthesis,
title = {Mainframe footprint and consumption, an evaluation based on sustainable IT approach},
author = {Pierre Jacquet},
school = {Université Paris 12},
year = {2021},
url = {https://jacquetpi.github.io/pdf/mainframe-consumption.pdf},
}
