On Coordinated Scheduling of Radio and Computing Resources in Cloud-RAN

Abstract

Cloud Radio Access Network is a promising mobile network architecture based on centralizing the baseband processing of many cellular base stations in a BBU (BaseBand Unit) pool. Such architecture has many advantages. However, computing resources are shared among the base stations connected to the BBU pool. It is challenging to schedule the processing of users' data, especially on overloaded BBU pools, while respecting the time constraints imposed by the Hybrid Automatic Repeat Request (HARQ) mechanism. Given that the processing time of users' data and the computing requirement depends on the radio parameters such as the Modulation and Coding Scheme (MCS), we propose to enable the coordination between radio and computing resources schedulers; such coordination makes the selection of MCS dependent on the availability of radio and computing resources and on the ability to process data while respecting the HARQ-deadline. In this context, we propose and evaluate three Integer Linear Programming (ILP)-based schemes and three low-complexity heuristics, demonstrating their ability to reduce the wasted transmission power. Moreover, we evaluate the performance of the coordination under a multiservices scenario consisting of two services having heterogeneous requirements, enhanced Mobile Broadband (eMBB) and Ultra-Reliable Low-Latency Communication (URLLC).

Policy-Gradient-Based Reinforcement Learning for Computing Resources Allocation in O-RAN

“…In our previous works [13] [14], we have considered computing resource allocation in Cloud-RAN architecture. Considering the problem of limited computing resources where frames compete for processing resources, [13] introduces Integer Linear Programming models that maximize throughput and fairness allocation.…”

“…In our previous works [13] [14], we have considered computing resource allocation in Cloud-RAN architecture. Considering the problem of limited computing resources where frames compete for processing resources, [13] introduces Integer Linear Programming models that maximize throughput and fairness allocation. The models permit coordinating between the radio and computing schedulers such that the radio parameter Modulation and Coding Scheme (MCS) index can be modified at the request of the computing scheduler.…”

“…For a given frame, the MCS index could be decreased and would lead to reduced throughput, but the computing scheduler would be able to allocate the required computing resources to process this frame. Knowing that solving an ILP problem is NP-Hard, [14] presents a Recurrent Neural Network Model that aims at performing as close as possible to the ILP model presented in [13] but with lower complexity.…”

“…Hence these O-DUs could fail to execute the decoding functions for all users. Given that there is a deadline for execution equal to 2 ms due to HARQ [13], some of the O-DUs will fail to process some of their users' frames, and this would trigger the HARQ mechanism; the frames will be retransmitted, and radio resources will be reserved for the future retransmission. We aim to use RL to learn which users should be admitted and which should halt their transmission.…”

“…The number of RBs per eMBB user follows a uniform random distribution ranging from 20 to 40 RBs, and the range is 1 to 5 for URLLC users. The deadline for eMBB traffic is 2ms as in [13], and the deadline for URLLC is 0.25ms. The number of TTI over which we aim to optimize fairness transmissions Algorithm 1: RL algorithm The required processing time is provided in [20] as a function of the MCS, the number of RBs, and the CPU frequency.…”