10ème Journéee Thématique du groupe COSMOS

Cette journée aura lieu à PARIS, le 02 décembre 2025.

La participation à la journée est libre mais l'inscription est obligatoire pour des soucis d'organisation. Le lien d'inscription est en haut de cette page.

Lieu :  Salle 534
Tour 26-00
Université Jussieu, 
4 place Jussieu 75005 Paris
Métro Jussieu. 

Orateurs : Elen Anton,  Stéphan Plassard, Jonatha Anselmi, Thomas Hira, Mohamed-Harith Ibrahim, Farah Ait Salaht. 

Program:

9h30 Accueil

10h Elen Anton : "Comparing the performance of the redundancy-d models under the FCFS and ROS scheduling policies".

10h45 Thomas Hira : "Scheduling with non-observable, time-varying environments"

11h30 Stephan Plassart : "A learning algorithm of energy efficient speed schedules in real time systems".

12h15 Discussion sur le GDT

12h30 Repas

14h00 Jonatha Anselmi : "Adaptive Autoscaling in Serverless Platforms through Non-Stationary Gradient Descent".

14h45 Farah Ait Salaht : Réplication et placement optimisés des applications orienté service dans le Computing Computing

15h30 Pause

16h00 Mohamed-Harith Ibrahim : "Gestion adaptative des systèmes énergétiques par apprentissage par renforcement en contexte incertain".

16h45 Fin de la journée

Exposés : 

Mohamed-Harith Ibrahim : "Gestion adaptative des systèmes énergétiques par apprentissage par renforcement en contexte incertain".

Elen Anton : "Comparing the performance of the redundancy-d models under the FCFS and ROS scheduling policies".
Resumé:  In this talk, we analyse the response time of redundancy-d systems under the first-come first-served (FCFS) and random order of service (ROS) scheduling policies. Redundancy has received considerable attention as a dispatching paradigm that promises the potential for significant response time improvements. Previous work on redundancy-d models – with exponentially distributed service times and independent and identically distributed copies – has focused on response time results under FCFS or stability conditions under a wider range of scheduling policies, including ROS, and shows that the stability region coincides under both scheduling policies. We provide the first proof that the FCFS scheduling policy improves the mean response time of the redundancy-d model compared to the ROS policy. To do so, we present a novel coupling technique that exploits the knowledge of the steady-state distribution under the FCFS policy.

Stephan Plassart : "A learning algorithm of energy efficient speed schedules in real time systems".
Resumé:  We present an algorithm that learns the statistics of jobs arriving in a hard real-time system (HRTS), and uses them to estimate the optimal processor speed policy to execute online the real-time jobs. In this (HRTS), the job features (inter-arrival time, size, and deadline) are not known in advance, not even their probability distributions, but only revealed when the job is released. Optimal means that each job completes its execution before its deadline (hard real-time constraint) and the expected energy consumption is minimized. The learning mechanism consists of two successive phases. The first phase is the online measurement of the features of the incoming real-time jobs (release times, sizes, and deadlines). These are unknown at the beginning of the learning period. The second phase is the computation of the optimal processor speed, to be used during the application phase to minimize the energy consumption under the constraint that each job completes before its deadline. We show that this approach is very efficient in practice because its regret remains bounded with a high probability. Experiments show that the energy consumption of the learned speed policy is very close to the performance of the optimal speed policy computed offline using the full knowledge of the job features.

Jonatha Anselmi : "Adaptive Autoscaling in Serverless Platforms through Non-Stationary Gradient Descent".
Resumé: As the adoption of serverless computing platforms continue to grow, designing autoscaling policies that strike the right balance between energy efficiency and user-perceived performance has become a central challenge. In this talk, we propose an online learning algorithm with theoretical convergence guarantees that dynamically tunes control parameters in a serverless autoscaling environment.  The proposed algorithm, grounded in stochastic gradient descent, learns online the optimal values of three key control parameters: (i) the target stock size of pre-warmed (idle) functions, (ii) the threshold triggering provisioning actions, and (iii) the expiration rate of idle resources. We prove that, under Markovian dynamics, the algorithm converges to the parameter set that minimizes a cost function capturing the tradeoff between energy consumption and response latency. In addition, we demonstrate that its structure naturally supports parallelization, significantly accelerating convergence. Extensive numerical experiments show that our method outperforms existing baselines, including recent deep learning-based approaches, even under  non-Markovian settings—highlighting both its robustness and practical viability for next-generation serverless infrastructures.

Thomas Hira : "Scheduling with non-observable, time-varying environments".
Resumé:  Non-observability of the state of environments, gives rise to interesting problems in scheduling. In this talk, we explore, in this scheduling context, what is possible under three levels of knowledge about the underlying Markovian dynamics. First, when the parameters are fully known, we present a belief-state based approach and illustrate it on a simple yet instructive two-state model. Second, with no prior knowledge of the Markov chains, we identify fundamental limits faced by any learning agent and propose a simple method that performs as well as the information allows. Finally, we show that even a small amount of prior knowledge, like a bound on the bias span, can yield substantial improvements.

Farah Ait Salaht : Réplication et placement optimisés des applications orienté service dans le Computing Computing
Résumé : Les applications modernes, structurées en graphes de microservices et déployées le long du computing continuum (de l'edge au cloud), doivent traiter en continu des flux de données très variables tout en maintenant une faible latence et une qualité de service élevée. Cette présentation introduit un modèle d'optimisation conjointe de la réplication et du placement des microservices, formulé comme un problème de satisfaction de contraintes (CSP). Le modèle qui se veut générique et facile à mettre à niveau capture à la fois les exigences des applications (latence, ressources) et l'hétérogénéité des infrastructures (capacités de calcul, stockage et réseau) pour déterminer automatiquement le degré de réplication et l'emplacement des microservices en fonction de la charge.