RT Conference Proceedings
T1 vrAIn: a deep learning approach tailoring computing and radio resources in virtualized RANs
A1 Ayala Romero, José A.
A1 Garcia Saavedra, Andres
A1 Gramaglia, Marco
A1 Costa-Pérez, Xavier
A1 Banchs Roca, Albert
A1 Alcaraz, Juan J.
AB The virtualization of radio access networks (vRAN) is thelast milestone in the NFV revolution. However, the complexdependencies between computing and radio resources makevRAN resource control particularly daunting. We presentvrAIn, a dynamic resource controller for vRANs based ondeep reinforcement learning. First, we use an autoencoderto project high-dimensional context data (traffic and signalquality patterns) into a latent representation. Then, we use adeep deterministic policy gradient (DDPG) algorithm basedon an actor-critic neural network structure and a classifierto map (encoded) contexts into resource control decisions.We have implemented vrAIn using an open-source LTEstack over different platforms. Our results show that vrAInsuccessfully derives appropriate compute and radio controlactions irrespective of the platform and context: (i) it providessavings in computational capacity of up to 30% overCPU-unaware methods; (ii) it improves the probability ofmeeting QoS targets by 25% over static allocation policiesusing similar CPU resources in average; (iii) upon CPU capacityshortage, it improves throughput performance by 25%over state-of-the-art schemes; and (iv) it performs close to optimalpolicies resulting from an offline oracle. To the best ofour knowledge, this is the first work that thoroughly studiesthe computational behavior of vRANs, and the first approachto a model-free solution that does not need to assume anyparticular vRAN platform or system conditions.
PB Association for Computing Machinery
SN 978-1-4503-6169-9
YR 2019
FD 2019-09-05
LK https://hdl.handle.net/10016/29093
UL https://hdl.handle.net/10016/29093
LA eng
NO Proceeding of: 25th Annual International Conference on Mobile Computing and Networking (MobiCom'19), October 21-25, 2019, Los Cabos, Mexico.
NO The work ofUniversity Carlos III of Madrid was supported by H2020 5GMoNArchproject (grant agreement no. 761445) and H20205G-TOURS project (grant agreement no. 856950). The workof NEC Laboratories Europe was supported by H2020 5GTRANSFORMERproject (grant agreement no. 761536) and5GROWTH project (grant agreement no. 856709). The workof University of Cartagena was supported by Grant AEI/FEDERTEC2016-76465-C2-1-R (AIM) and Grant FPU14/03701.
DS e-Archivo
RD 7 jul. 2024