4 présentations

• 
  15h30 - 15h55

  Network Design Problem with Vulnerability and Budgeting Constraints

  • Jai Kumar Drave, prés., HEC Montreal
  • Yossiri Adulyasak, HEC Montréal
  • Okan Arslan, GERAD, HEC Montréal

  The network design problem with vulnerability and budgeting constraints (NDPVBC) is a generalization of the NDPVC, in which we consider the tradeoff between network survivability and quality of service subject to a budget constraint. Given a network and a set of origin-destination pairs with communication needs, the NDPVC selects a subset of the edges at minimum cost in order to satisfy the demand at a certain quality of service while ensuring survivability of the network. The quality of service is expressed as a function of the number of edges used to connect the origins to their destinations, while the survivability is a function of the number of simultaneous edge failures in the network. The selected edges must provide a primary path for every demand, and a backup path when some edges fail in the network. In the literature, the quality of service and the survivability are considered to be parameters. We relax this assumption, consider them as variables of the problem and investigate the tradeoff between them subject to a budget constraint. We present a mathematical model for solving the problem, and discuss the results of a computational study and our findings.

• 
  15h55 - 16h20

  Multilayer route and spectrum assignment problems in elastic optical networks

  • Ehsan Rezagholizadeh, prés.,

  we propose a route and spectrum allocation (RSA) algorithm for the elastic optical network (EON). We consider a linear model for OTN as the electrical layer, and EON as the optical layer. We use a multilayer auxiliary graph and manipulate the edge cost adaptively to apply different policies. Using the layer-one processing power and software-defined network as a centralized controller empowers the proposed algorithm to gain more spectrum utilization with awareness of the spectrum continuity and contiguity constraints. The proposed heuristic method improves the performance of the network, as well as reduces the blocking probability.

• 
  16h20 - 16h45

  Distributed Task Offloading and Service Caching in Cloudlet Computing Networks

  • Mohammad Reza Golzari Oskoui, prés., Ecole Polytechnique de Montreal
  • Brunilde Sansò, GERAD, Polytechnique Montréal

  Cloudlet computing is a distributed computational paradigm aimed at reducing latency by bringing computation and data storage closer to data sources. However, modern user applications consist of interdependent tasks, complicating the task offloading problem. Moreover, as each task necessitates a pre-loaded service on the server to execute, determining which services to cache at cloudlets adds another layer of complexity. To tackle these issues, we present a framework comprising users, cloudlets, and a broker. Our distributed approach employs deep reinforcement learning and transfer learning to efficiently handle task offloading and service caching. The method defines states for the learning agent to understand the environment in terms of task dependency levels and the services cached by cloudlets. Furthermore, by introducing a broker in the framework and performing statistical computations to rank the required services, we address the service caching problem. In order to accelerate learning speed among users, the broker updates its neural network based on the cached services on cloudlets. It then broadcasts the learned coefficients to users for leveraging transfer learning effectively. Through simulation, our algorithm outperforms existing benchmarks by optimizing cloudlet selection for task offloading and service caching, thereby reducing application completion time.

• 
  16h45 - 17h10

  Network design with vulnerability constraints and probabilistic edge reliability

  • Okan Arslan, prés., GERAD, HEC Montréal
  • Gilbert Laporte, HEC Montréal

  The Network design problem with vulnerability constraints and probabilistic edge reliability (NDPVC-PER) is an extension of the NDPVC obtained by additionally considering edge reliability. We consider the design of a telecommunication network in which every origin-destination pair is connected by a hop-constrained primal path, and by a hop-constrained backup path when certain edges in the network fail. The edge failures occur with respect to their reliability, and the network is designed by considering a minimum reliability level. Therefore, an hop-constrained backup path must be built by considering all simultaneous edge failures that have a certain probability of realization. While there exist models to solve the NDPVC without enumerating all edge subsets, edge reliability cannot be dealt with by applying the techniques applied to the NDPVC. Therefore, we develop models based on a new concept of resilient length-bounded cuts, and solve the NDPVC-PER without edge set enumerations. We perform extensive testing of the model to determine the best performing settings, and demonstrate the computational efficiency of the developed model. Our findings on these instances show that, in the dataset considered in this study, increasing the reliability level from 90% to 95% increases the average cost only by 12.4%, while increasing it from 95% to 99% level yields a cost increase of 93.9%.