Events in September–October 2022
-
Talk Anshul Gandhi: "Resource Management and Scheduling in Today’s Cloud"
Talk Anshul Gandhi: "Resource Management and Scheduling in Today’s Cloud"
–
September 22, 2022Title: Resource Management and Scheduling in Today’s CloudAbstract: The cloud computing landscape today is evolving quickly. New computing and deployment paradigms, like serverless computing and microservices architecture, are augmenting the classical VM-based cloud that we are all familiar with. In this new cloud, the basic question still remains "How do we efficiently utilize available resources to maintain low application latency". In this talk, I will discuss this evolving landscape briefly and then describe some recent works from our lab that address resource management and scheduling problems in the context. I will start by discussing ENSURE, our serverless scheduling solution that aims to maintain serverless function latency while minimizing the amount of resources needed. The specific challenges here are the diversity in serverless workload characteristics and the cold start penalty that non-warm containers incur when launched. I will then discuss our analytical and systems contributions in the context of colocating workload in the cloud with the goal of minimizing performance interference while maximing resource usage. I will end with a brief discussion of other projects in the cloud computing space we have looked at, including deploying Spark in an NDP environment, optimizing TensorFlow DNN training times in multi-GPU servers, and tuning the configuration of microservices architecture applications.Bio: Anshul Gandhi is an Associate Professor in the Computer Science Department at Stony Brook University. He received his Ph.D. from Carnegie Mellon University in 2013 and then spent a year as a postdoc at the IBM T. J. Watson Research Center. His current research focuses on performance modeling in distributed systems, and is funded by an NSF Career award, an IBM Faculty award, and a Google Research award. His contributions to performance modeling were recently recognized by an ACM Sigmetrics Rising Star Award.Bâtiment IMAG (442) -
Seminar Joao Comba: Data Visualization for The Understanding of COVID-19
Seminar Joao Comba: Data Visualization for The Understanding of COVID-19
–
October 12, 2022Title: Data Visualization for The Understanding of COVID-19
Speaker: Joao Comba (UFRGS)
Abstract: In this talk, I will describe several data visualization
projects related to COVID-19 data analysis. The first work describes
an analysis tool for comparing the evolution of COVID-19
data in different regions worldwide. The main component of our tool is
a similarity engine that searches for similar patterns of COVID-19 time
series and a storytelling dashboard that identifies the waves of COVID-
19. We also introduce a new type of difference plot that allows a more
straightforward comparison of time series evolution. The second work
leverages the big-data infrastructure we developed for real-time
analysis to study two years of clinical admissions in the public health
system of Brazil, composed of more than 100 million records. The third
work is related to the diagnosis of COVID-19 from computed tomography
and clinical data. We will describe COVID-VR, a novel approach for
classifying pulmonary diseases based on volume rendering images of the
lungs taken from different angles, thus providing a global view of the
entire lung in each image. I will conclude the talk with a summary of
my current research projects.
Short-bio: João L. D. Comba is a Full Professor at the Instituto de
Informática at UFRGS, where he conducts research activities,
undergraduate and graduate education, and supervises doctoral,
master, and scientific initiation students. He holds a Ph.D. in
Computer Science from Stanford University, United States. 2000. His
research areas include data visualization and analysis, data science,
geometric algorithms, and spatial data structures. He is currently co-
chair of the Visualization Corner of the Journal Computing in Science &
Engineering and associate editor of the Computer & Graphics Journal.Bâtiment IMAG (442) -
Séminaire Antoine Oustry: Polynomial optimization for parameterizing dynamical systems
Séminaire Antoine Oustry: Polynomial optimization for parameterizing dynamical systems
–
October 13, 2022Polynomial optimization for parameterizing dynamical systems
Abstract: We consider a polynomial dynamical system that depends on some tunable parameters. We raise the question of how to configure this system optimally, in terms of costs and stability.
More precisely, we will give an overview of polynomial optimization approaches to
(a) find a (stable) equilibrium point of the dynamical system that minimizes a polynomial function
(b) approximate regions of stability around the equilibrium points.
We illustrate our work with some applications in power engineering. -
Séminaire Philippe Swartvagher (Inria Bordeaux): Possible interactions between task-based runtime systems and communication libraries
Séminaire Philippe Swartvagher (Inria Bordeaux): Possible interactions between task-based runtime systems and communication libraries
–
October 20, 2022 -
Seminar Bryce Ferguson: "Information and Influence: Overcoming and Exploiting Uncertainty in Congestion Games"
Seminar Bryce Ferguson: "Information and Influence: Overcoming and Exploiting Uncertainty in Congestion Games"
–
October 27, 2022Bryce Ferguson (https://web.ece.ucsb.edu/~blf/)Title: Information and Influence: Overcoming and Exploiting Uncertainty in Congestion GamesAbstract: In large-scale, socio-technical systems (such as traffic networks, power grids, supply chains, etc.) the operating efficiency depends heavily on the actions of human users. It is well known that when users act in their own self-interest, system performance can be sub-optimal. Our capabilities in alleviating this inefficiency rely on our knowledge of user decision making and various system parameters. In this talk, I will present two settings where information affects our ability to influence system performance. In the first, we consider designing monetary incentives for players in a congestion game without exact knowledge of users’ price-sensitivity or path latency-characteristics; we provide a comparison of the effectiveness of different incentive types and quantify the value of different pieces of information. In the second, we consider a flipped paradigm, where the system operator has more information about system parameters than the users and can selectively reveal pertinent information. We show, in the context of Bayesian-congestion games, that signaling information to users has the opportunity to improve system performance but also the capability to make performance worse than if no information were shared at all. We then show the advantages of concurrently using monetary incentives and information signals by providing bounds on the benefit to system performance and methods to find optimal mechanisms of each type.Bâtiment IMAG (406)Saint-Martin-d'Hères, 38400France
