Camille Coti
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Camille Coti has a total of 79 co-authors. Here are the top co-authors by number of shared publications:
Thomas Herault
5 shared publications- Laure Petrucci
Université Sorbonne Paris Nord, LIPN, CNRS UMR 7030
4 shared publications Franck Butelle
3 shared publicationsFranck Cappello
3 shared publications- Vladimir Reinharz
Université du Québec à Montréal
3 shared publications - Wilfried Agbeto
Université du Québec à Montréal
3 shared publications
Latest Publications
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 624-640
Untangling GPU Power Consumption: Job-Level Inference in Cloud Shared Settings
As the demand for AI-driven workloads increases, the energy consumption of Graphics Processing Units (GPUs) devices has come under intense scrutiny, particularly in hyperscale data centers where large numbers of accelerators are centralized and leased to diverse clients. In the context of cloud hyperscalers, GPUs power monitoring presents several challenges that vary depending on the product offered. The monitoring capabilities of physical devices may be limited or even absent for some products. However, given the substantial energy demands of GPUs, power monitoring is essential for both cloud providers and clients. Operators require tools to manage power distribution effectively, such as balancing workloads across Power Distribution Units (PDUs), while clients need visibility into power usage to optimize their workloads for energy efficiency. To address these challenges, we propose methods for estimating the energy consumption of jobs running on GPU devices in cloud environments, spanning from shared and managed offerings like ML-as-a-Service (MLaaS) to less managed products (e.g., Infrastructure-as-a-Service (IaaS)). 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.
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 624-640
Open AccessUntangling GPU Power Consumption: Job-Level Inference in Cloud Shared Settings
As the demand for AI-driven workloads increases, the energy consumption of Graphics Processing Units (GPUs) devices has come under intense scrutiny, particularly in hyperscale data centers where large numbers of accelerators are centralized and leased to diverse clients. In the context of cloud hyperscalers, GPUs power monitoring presents several challenges that vary depending on the product offered. The monitoring capabilities of physical devices may be limited or even absent for some products. However, given the substantial energy demands of GPUs, power monitoring is essential for both cloud providers and clients. Operators require tools to manage power distribution effectively, such as balancing workloads across Power Distribution Units (PDUs), while clients need visibility into power usage to optimize their workloads for energy efficiency. To address these challenges, we propose methods for estimating the energy consumption of jobs running on GPU devices in cloud environments, spanning from shared and managed offerings like ML-as-a-Service (MLaaS) to less managed products (e.g., Infrastructure-as-a-Service (IaaS)). 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.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (2) , 822-833
Cinergy: Deterministic Power Monitoring for Carbon Accounting in the Cloud
Cloud providers now commonly offer tools to monitor the environmental impact of hosted services, aiming to inform customers about the carbon footprint of their infrastructures. A significant share of this impact 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. This article proposes a framework, called Cinergy, to standardize how cloud providers integrate the power consumption of virtual resources into their carbon calculators. It defines a baseline to deterministically assess the power consumption of a provisioned Virtual Machine (VM)—i.e., the same usage should yield the same power estimate—while still accounting for consolidation gains enabled by virtualization. Our approach, constructed and evaluated empirically, 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, with differences reaching 85% for the same VM size. We also show that quantity-based carbon accounting methodologies can drastically underestimate actual CO2 emissions-by up to a factor of three.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cloud Computing
- Vol. 14 (2) , 822-833
Cinergy: Deterministic Power Monitoring for Carbon Accounting in the Cloud
Cloud providers now commonly offer tools to monitor the environmental impact of hosted services, aiming to inform customers about the carbon footprint of their infrastructures. A significant share of this impact 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. This article proposes a framework, called Cinergy, to standardize how cloud providers integrate the power consumption of virtual resources into their carbon calculators. It defines a baseline to deterministically assess the power consumption of a provisioned Virtual Machine (VM)—i.e., the same usage should yield the same power estimate—while still accounting for consolidation gains enabled by virtualization. Our approach, constructed and evaluated empirically, 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, with differences reaching 85% for the same VM size. We also show that quantity-based carbon accounting methodologies can drastically underestimate actual CO2 emissions-by up to a factor of three.
- 1 April 2025
- Published by Springer Nature
- p. 131-145
Parallel Maximal Common Subgraphs with Labels for Molecular Biology
- 1 April 2025
- Published by Springer Nature
- p. 131-145
Parallel Maximal Common Subgraphs with Labels for Molecular Biology
- Open Access
FASTiso: Fast Algorithm on Search state Tree for subgraph ISOmorphism in graphs of any size and density
Subgraph isomorphism is a fundamental combinatorial problem that involves finding one or more occurrences of a pattern graph within a target graph. It arises in a wide range of application domains, including biology, chemistry, social network analysis, and pattern recognition. Although subgraph isomorphism is NP-complete in the general case, many exact algorithms allow it to be solved in practice on many instances. However, the increasing size and structural diversity of graph datasets continue to pose significant challenges in terms of robustness and scalability. In this article, we propose FASTiso, an exact subgraph isomorphism algorithm that emphasizes a strong consistency between the variable ordering strategy and the pruning rules used during search. This design enables a unified exploitation of structural information throughout the exploration process, leading to improved efficiency and stable performance across heterogeneous graph structures. An extensive experimental evaluation on widely used synthetic and real-world benchmarks shows that FASTiso consistently outperforms reference solvers such as VF3, VF3L, and RI, and achieves competitive performance compared to constraint programming–based approaches (Glasgow, PathLad+), while outperforming them on most datasets. The results further demonstrate that FASTiso remains highly efficient on small instances and scales well to large graphs, while maintaining a lower memory footprint than most evaluated solvers. The peak memory usage is 7.74 GB for FASTiso, 36.19 GB for PathLad+, over 500 GB for Glasgow, 9.62 GB for VF3/VF3L, and 4.31 GB for RI. FASTiso code is available at https://gitlab.info.uqam.ca/cbe/fastiso as a C++ implementation, a Python module, and an integration within an extended version of NetworkX. The implementations support simple graphs and multigraphs, directed or undirected, with labels on nodes, edges, or both.
Open AccessFASTiso: Fast Algorithm on Search state Tree for subgraph ISOmorphism in graphs of any size and density
Subgraph isomorphism is a fundamental combinatorial problem that involves finding one or more occurrences of a pattern graph within a target graph. It arises in a wide range of application domains, including biology, chemistry, social network analysis, and pattern recognition. Although subgraph isomorphism is NP-complete in the general case, many exact algorithms allow it to be solved in practice on many instances. However, the increasing size and structural diversity of graph datasets continue to pose significant challenges in terms of robustness and scalability. In this article, we propose FASTiso, an exact subgraph isomorphism algorithm that emphasizes a strong consistency between the variable ordering strategy and the pruning rules used during search. This design enables a unified exploitation of structural information throughout the exploration process, leading to improved efficiency and stable performance across heterogeneous graph structures. An extensive experimental evaluation on widely used synthetic and real-world benchmarks shows that FASTiso consistently outperforms reference solvers such as VF3, VF3L, and RI, and achieves competitive performance compared to constraint programming–based approaches (Glasgow, PathLad+), while outperforming them on most datasets. The results further demonstrate that FASTiso remains highly efficient on small instances and scales well to large graphs, while maintaining a lower memory footprint than most evaluated solvers. The peak memory usage is 7.74 GB for FASTiso, 36.19 GB for PathLad+, over 500 GB for Glasgow, 9.62 GB for VF3/VF3L, and 4.31 GB for RI. FASTiso code is available at https://gitlab.info.uqam.ca/cbe/fastiso as a C++ implementation, a Python module, and an integration within an extended version of NetworkX. The implementations support simple graphs and multigraphs, directed or undirected, with labels on nodes, edges, or both.
- Open Access
Parallel maximal common subgraphs with labels for molecular biology
Advances in graph algorithmics have allowed in-depth study of many natural objects from molecular biology or chemistry to social networks. Particularly in molecular biology and cheminformatics, understanding complex structures by identifying conserved sub-structures is a key milestone towards the artificial design of novel components with specific functions. Given a dataset of structures, we are interested in identifying all maximum common connected partial subgraphs between each pair of graphs, a task notoriously NP-Hard. In this work, we present parallel algorithms over shared and distributed memory to enumerate all maximal connected common sub-graphs between pairs of arbitrary multi-directed graphs with labels on their edges. We offer an implementation of these methods and evaluate their performance on the non-redundant dataset of all known RNA 3D structures. We show that we can compute the exact results in a reasonable time for each pairwise comparison while taking into account a much more diverse set of interactions—resulting in much denser graphs—resulting in an order of magnitude more conserved modules. All code is available at https://gitlab.info.uqam.ca/cbe/pasigraph and results in the branch results .
Open AccessParallel maximal common subgraphs with labels for molecular biology
Advances in graph algorithmics have allowed in-depth study of many natural objects from molecular biology or chemistry to social networks. Particularly in molecular biology and cheminformatics, understanding complex structures by identifying conserved sub-structures is a key milestone towards the artificial design of novel components with specific functions. Given a dataset of structures, we are interested in identifying all maximum common connected partial subgraphs between each pair of graphs, a task notoriously NP-Hard. In this work, we present parallel algorithms over shared and distributed memory to enumerate all maximal connected common sub-graphs between pairs of arbitrary multi-directed graphs with labels on their edges. We offer an implementation of these methods and evaluate their performance on the non-redundant dataset of all known RNA 3D structures. We show that we can compute the exact results in a reasonable time for each pairwise comparison while taking into account a much more diverse set of interactions—resulting in much denser graphs—resulting in an order of magnitude more conserved modules. All code is available at https://gitlab.info.uqam.ca/cbe/pasigraph and results in the branch results .
- Open Access
Parallel Computation of Multi-Slice Clustering of Third-Order Tensors
Machine Learning approaches like clustering methods deal with massive datasets that present an increasing challenge. We devise parallel algorithms to compute the Multi-Slice Clustering (MSC) for 3rd-order tensors. The MSC method is based on spectral analysis of the tensor slices and works independently on each tensor mode. Such features fit well in the parallel paradigm via a distributed memory system. We show that our parallel scheme outperforms sequential computing and allows for the scalability of the MSC method.
Open AccessParallel Computation of Multi-Slice Clustering of Third-Order Tensors
Machine Learning approaches like clustering methods deal with massive datasets that present an increasing challenge. We devise parallel algorithms to compute the Multi-Slice Clustering (MSC) for 3rd-order tensors. The MSC method is based on spectral analysis of the tensor slices and works independently on each tensor mode. Such features fit well in the parallel paradigm via a distributed memory system. We show that our parallel scheme outperforms sequential computing and allows for the scalability of the MSC method.
- 20 May 2022
- Published by Springer Nature
- p. 63-80
SKaMPI-OpenSHMEM: Measuring OpenSHMEM Communication Routines
- 20 May 2022
- Published by Springer Nature
- p. 63-80
SKaMPI-OpenSHMEM: Measuring OpenSHMEM Communication Routines
- 4 February 2021
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 33 (11) , e6179
DiPOSH: A portable OpenSHMEM implementation for short API‐to‐network path
Summary: In this article, we introduce DiPOSH, a multi‐network, distributed implementation of the OpenSHMEM standard. The core idea behind DiPOSH is to have an API‐to‐network software stack as slim as possible, in order to minimize the software overhead. Following the heritage of its non‐distributed parent POSH, DiPOSH's communication engine is organized around the processes' shared heaps, and remote communications are moving data from and to these shared heaps directly. This article presents its architecture and several communication drivers, including one that takes advantage of a helper process, called the Hub, for inter‐process communications. This architecture allows use to explore different options for implementing the communication drivers, from using high‐level, portable, optimized libraries to low‐level, close to the hardware communication routines. We present the perspectives opened by this additional component in terms of communication scheduling between and on the nodes. DiPOSH is available athttps://github.com/coti/DiPOSH.
- 4 February 2021
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 33 (11) , e6179
DiPOSH: A portable OpenSHMEM implementation for short API‐to‐network path
Summary: In this article, we introduce DiPOSH, a multi‐network, distributed implementation of the OpenSHMEM standard. The core idea behind DiPOSH is to have an API‐to‐network software stack as slim as possible, in order to minimize the software overhead. Following the heritage of its non‐distributed parent POSH, DiPOSH's communication engine is organized around the processes' shared heaps, and remote communications are moving data from and to these shared heaps directly. This article presents its architecture and several communication drivers, including one that takes advantage of a helper process, called the Hub, for inter‐process communications. This architecture allows use to explore different options for implementing the communication drivers, from using high‐level, portable, optimized libraries to low‐level, close to the hardware communication routines. We present the perspectives opened by this additional component in terms of communication scheduling between and on the nodes. DiPOSH is available athttps://github.com/coti/DiPOSH.
- 12 October 2020
- Published by Springer Nature in Formal Methods in System Design
- Vol. 57 (1) , 3-33
Quasi-optimal partial order reduction
- 12 October 2020
- Published by Springer Nature in Formal Methods in System Design
- Vol. 57 (1) , 3-33
Quasi-optimal partial order reduction