|Speaker :||Andrea Araldo|
|Télécom Paris Sud|
|Time:||2:00 pm - 3:00 pm|
Edge Computing (EC) is a network paradigm in which computational resources, e.g., CPU, memory, are distributed in the access networks, close to the final users. Network elements at the edge, e.g., base stations, road side units, access points, become intelligent and can be used by multiple third party Service Providers (SPs) to run (a part of) their application at the edge. Such applications can be heterogeneous, ranging from video-streaming to augmented reality, to services needed to autonomous driving.
In order to enable this, the Network Operator (NO), owning the physical infrastructure, needs to virtualize edge resources and allocate them to the SPs, which represent the tenants.
Unlike Cloud Computing, EC resources are scarce and contention arises between SPs. Therefore, the NO needs to carefully allocate virtual resources among them, in order to increase its utility, e.g., reduction in bandwidth consumption, user QoE. Such allocation is challenging, since SPs must be kept isolated, i.e., data and computation of each SP must be inaccessible to other SPs and to the NO. In other words, SPs are black boxes for the NO, which still must be able to properly allocate resources among them.
A. Araldo will present some solutions to the problem of Multi-Tenant Resource Allocation in EC, described above, based on two approaches: explicit information exchange between SPs and NO or data-driven allocation algorithms, as reinforcement learning and stochastic perturbation.
Finally, A. Araldo will present some preliminary results related to the performance of large scale augmented reality applications implemented via a distributed system of Edge AI boards, which are small and cheap embedded devices equipped with GPU, able to run deep neural networks.