INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Untangling GPU Power Consumption: Job-Level Inference in Cloud Shared Settings
No full textInternational audienceAs 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
Supervised aggregation of anomaly score functions for active anomaly detection
No full textInternational audienceDetecting rare anomalies in batches of multidimensional data is challenging. We propose a supervised active-learning framework that sends a small number of data points from each batch to an expert for labeling as 'anomaly' or 'nominal', via two mechanisms: (i) points most likely to be an anomaly in the eyes of a supervised classifier trained on previously-labeled data; and (ii) points suggested by an active learner. Instead of, however, training the supervised classifier directly on the current set of labeled raw data points, we treat the scores calculated by an ensemble of M unsupervised anomaly detectors on each data point as if they were the learner's input features. This approach generalizes earlier attempts to linearly aggregate unsupervised anomaly detector scores, and broadens the scope of such methods to ordered data like time series. Results suggest that this method usually outperforms-often significantly-linear strategies. The Python library acanag provides an implementation of the proposed method
Autoregressive Multiplier Bootstrap for In-situ Error Estimation and Quality Monitoring of Finite Time Averages in Turbulent Flow Simulations
No full textInternational audienceIn Computational Fluid Dynamics (CFD), and particularly within Direct Numerical Simulation (DNS) and Large Eddy Simulation (LES), the computational cost is largely dictated by the effort required to obtain statistically converged quantities such as time-averaged fields and higher-order moments. Despite the importance of accurately quantifying statistical uncertainty in unsteady simulations, no continuous and cost-effective, on-line method currently exists for monitoring the convergence quality of such statistics during runtime. This work introduces a novel, fully on-line bootstrapping approach to estimate the variance of finite-time averages without requiring the estimation of the flows Auto-Correlation Function (ACF). Unlike existing methods that rely on ACF estimation, which are often impractical due to excessive storage demands in large-scale simulations, or require off-line processing or a priori modeling assumptions, our method operates entirely during the simulation and incurs minimal overhead. The proposed technique employs a recursive update of bootstrap replicates of the time average, using correlated random weights generated via an autoregressive model. This formulation is computationally efficient: the update cost scales linearly with the number of bootstrap replicates and the dimensionality of the flow field, and the autoregressive model is inexpensive to evaluate. The method only requires storage of a small number of fields, making it suitable for large-scale CFD applications. We demonstrate the effectiveness of the approach on synthetic data from the Ornstein-Uhlenbeck process and on two canonical LES cases: a turbulent pipe flow and a round jet. We further discuss the methods applicability to simulations with non-uniform time stepping, highlighting its flexibility and robustness
On design, analysis, and hybrid manufacturing of microstructured blade-like geometries
No full textInternational audienceWith the evolution of new manufacturing technologies such as multi-material 3D printing, one can think of new type of objects that consist of considerably less, yet heterogeneous, material, consequently being porous, lighter and cheaper, while having the very same functionality as the original object when manufactured from one single solid material. We aim at questioning five decades of traditional paradigms in geometric CAD and focus at new generation of CAD objects that are not solid, but contain heterogeneous free-form internal microstructures. We propose a unified manufacturing pipeline that involves all stages, namely design, optimization, manufacturing, and inspection of microstructured free-form geometries. We demonstrate our pipeline on an industrial test case of a blisk blade that sustains the desired pressure limits, yet requires significantly less material when compared to the solid counterpart
Les anagrammes de ce titre
No full textNational audienceCet article se propose de résoudre le problème de combinatoire suivant : étant donné un multiensemble de lettres (par exemple les lettres du titre de cet article) et un ensemble de mots (par exemple les mots de notre correcteur orthographique), combien y a-t-il d'anagrammes, c'est-à-dire de séquences de mots utilisant exactement les lettres données ? Pour y parvenir, nous proposons d'une part une structure de données efficace pour représenter des multiensembles et d'autre part un algorithme d'énumération des anagrammes. En moins de temps qu'il n'en faut pour lire cette première page, nous aurons compté le nombre d'anagrammes de son titre pour un dictionnaire de plus de cent mille mots
Decentralized Machine Learning with Centralized Performance Guarantees via Gibbs Algorithms
No full textComplexity Gaps between Point and Interval Temporal Graphs for some Reachability Problems
No full textTemporal graphs arise when modeling interactions that evolve over time. They usually come in several flavors, depending on the number of parameters used to describe the temporal aspects of the interactions: time of appearance, duration, delay of transmission. In the point model, edges appear at specific points in time, whereas in the more general interval model, edges can be present over specific time intervals. In both models, the delay for traversing an edge can change with each edge appearance. When time is discrete, the two models are equivalent in the sense that the presence of an edge during an interval is equivalent to a sequence of point-in-time occurrences of the edge. However, this transformation can drastically change the size of the input and has implications for complexity. Indeed, we show a gap between the two models with respect to the complexity of the classical problem of computing a fastest temporal path from a source vertex to a target vertex, i.e. a path where edges can be traversed one after another in time and such that the total duration from source to target is minimized. It can be solved in near-linear time in the point model, while we show that the interval model requires quadratic time under classical assumptions of fine-grained complexity. With respect to linear time, our lower bound implies a factor of the number of vertices, while the best known algorithm has a factor of the number of underlying edges. We also show a similar complexity gap for computing a shortest temporal path, i.e. a temporal path with a minimum number of edges. Here our lower bound matches known upper bounds up to a logarithmic factor. Interestingly, we show that near-linear time for fastest temporal path computation is possible in the interval model when it is restricted to uniform delay zero, i.e., when traversing an edge is instantaneous. However, this special case is not exempt from our lower bound for shortest temporal path computation. These two results should be contrasted with the computation of a foremost temporal path, i.e., a temporal path that arrives as early as possible. It is well known that this computation can be solved in near-linear time in both models.We also show that there is no gap in testing the all-to-all temporal connectivity of a temporal graph. We demonstrate a quadratic lower bound that applies to both the interval and point models and aligns with the existing upper bounds
An efficient construction of Raz's two-source randomness extractor with improved parameters
No full textRandomness extractors are algorithms that distill weak random sources into near-perfect random numbers. Two-source extractors enable this distillation process by combining two independent weak random sources. Raz's extractor (STOC '05) was the first to achieve this in a setting where one source has linear min-entropy (i.e., proportional to its length), while the other has only logarithmic min-entropy in its length. However, Raz's original construction is impractical due to a polynomial computation time of at least degree 4. Our work solves this problem by presenting an improved version of Raz's extractor with quasi-linear computation time, as well as a new analytic theorem with reduced entropy requirements. We provide comprehensive analytical and numerical comparisons of our construction with others in the literature, and we derive strong and quantum-proof versions of our efficient Raz extractor. Additionally, we offer an easy-to-use, open-source code implementation of the extractor and a numerical parameter calculation module
Optimal Fidelity Estimation from Binary Measurements for Discrete and Continuous Variable Systems
No full textInternational audienceEstimating the fidelity between a desired target quantum state and an actual prepared state is essential for assessing the success of experiments. For pure target states, we use functional representations that can be measured directly and determine the number of copies of the prepared state needed for fidelity estimation. In continuous variable (CV) systems, we use the Wigner function, which can be measured via displaced parity measurements. We provide upper and lower bounds on the sample complexity required for fidelity estimation, considering the worst-case scenario across all possible prepared states. For target states of particular interest, such as Fock and Gaussian states, we find that this sample complexity is characterized by the L 1 -norm of the Wigner function, a measure of Wigner negativity widely studied in the literature, in particular in resource theories of quantum computation. For discrete variable systems consisting of n qubits, we explore fidelity estimation protocols using Pauli string measurements. Similarly as for the CV approach, the sample complexity is shown to be characterized by the L 1 -norm of the characteristic function of the target state for both Haar random states and stabilizer states. Furthermore, in a general black box model, we prove that, for any target state, the optimal sample complexity for fidelity estimation is characterized by the smoothed L 1 -norm of the target state. To the best of our knowledge, this is the first time the L 1 -norm of the Wigner function provides a lower bound on the cost of some information processing task
Robust a posteriori estimation of probit-lognormal seismic fragility curves via sequential design of experiments and constrained reference prior
No full textInternational audienceA seismic fragility curve expresses the probability of failure of a structure conditional to an intensity measure (IM) derived from seismic signals. When only limited data is available, the practitioner often refers to the probit-lognormal model coupled with maximum likelihood estimation (MLE) to obtain estimates of these curves. This means that only a binary indicator of the state (BIS) of the structure is known, namely a failure or non-failure state indicator, when it is subjected to a seismic signal with an intensity measure IM. In this context, the objective of this work is to propose a method for optimally estimating such curves by obtaining the most precise estimate possible with the minimum of data. The novelty of our work is twofold. First, we present and show how to mitigate the likelihood degeneracy problem which is ubiquitous with small data sets and hampers frequentist approaches such as MLE. Second, we propose a novel strategy for sequential design of experiments (DoE) that selects seismic signals from a large database of synthetic or real signals via their IM values, to be applied to structures to evaluate the corresponding BISs. This strategy relies on a criterion based on information theory in a Bayesian framework. It therefore aims to sequentially designate the IM value such that the pair (IM, BIS) has on average, with respect to the BIS of the structure, the greatest impact on the posterior distribution of the fragility curve. The methodology is applied to a case study from the nuclear industry. The results demonstrate its ability to efficiently and robustly estimate the fragility curve, and to avoid degeneracy even with a limited amount of data, i.e., less than 100. Furthermore, we demonstrate that the estimates quickly reach the model bias induced by the probit-lognormal modeling. Eventually, two criteria are suggested to help the user stop the DoE algorithm