HARQ-aware allocation of computing resources in C-RAN

Abstract

The principal tenet of C-RAN is the softwarization of the base-band signal processing, which enables the sharing of computing resources among multiple radio heads. When the aggregate demand exceeds the processing capacity, a fraction of the radio packets is lost at PHY layer. Traditional computing resource allocation policies aim to minimize the packet loss rate.Dropping a PHY packet triggers a retransmission, unless the lost packet corresponds to the last available HARQ round, in which case the entirety of the radio resources spent on the multiple transmissions go to waste. This suggests that allocating computing resource accounting also for the HARQ transmission history may make a more efficient use of the bandwidth.We consider a simplified LTE uplink setting, and we measure the performance at the lower MAC layer (accuracy, goodput and average delay). We first compare the PHY-layer loss rate minimization and the cross-layer approaches using an ILP formulation. The cross-layer approach brings a tangible improvement, especially in accuracy. This suggests, for future work, that joint radio and computing resource allocation may further enhance spectral efficiency. We finally propose a probabilistic algorithm amenable to real-time operation which allows to mix strategies via parameter tuning, and we use it to explore the region of achievable goodput/accuracy trade-offs.

A Recurrent Neural Network Based Approach for Coordinating Radio and Computing Resources Allocation in Cloud-RAN

“…These two ILP problems were improved in [3]; in the latter, they gained an extra degree of freedom allowing them to modify the MCS index of users to better adapt the transmission parameters (i.e., MCS index) with the availability of computing resources. Unlike [10] where the performance is measured at the PHY layer, [11] takes into consideration the performance at the MAC layer. The different ILP algorithms have been compared with respect to different MAC layer metrics including goodput, bit-error-rate, and average delay.…”

“…The satisfaction ratio of user u ∈ Ur should be no more than 0.1 [11]. However, adjusting the MCS index to a lower one allows for decreasing the required processing time at the expense of reducing throughput.…”