Prescriptive Analytics for MEC Orchestration

Abstract

Orchestrating network and computing resources in Mobile Edge Computing (MEC) is an important item in the networking research agenda. In this paper, we propose a novel algorithmic approach to solve the problem of dynamically assigning base stations to MEC facilities, while taking into consideration multiple time-periods, and computing load switching and access latency costs. In particular, leveraging on an existing state of the art on mobile data analytics, we propose a methodology to integrate arbitrary time-period aggregation methods into a network optimization framework. We notably apply simple consecutive time period aggregation and agglomerative hierarchical clustering. Even if the aggregation and optimization methods represent techniques which are different in nature, and whose aim is partially overlapping, we show that they can be integrated in an efficient way. By simulation on real mobile cellular datasets, we show that, thanks to the clustering, we can scale with the number of time-periods considered, that our approach largely outperforms the case without time-period aggregations in terms of MEC access latency, and at which extent the use of clustering and time aggregation affects computing time and solution quality.

Complexity-Performance Trade-offs in Robust Access Point Clustering for Edge Computing

“…The main goal is to define MEC clusters as a set of BSs and users served by the same MEC host, so that at the end the whole area will be partitioned into MEC clusters, in order to offload the core network by maximizing intra MEC hosts communications. Similarly, in [8] where authors apply a temporal clustering model proposed in [31] on traffic demands of a real world dataset and integrate it into an orchestration model. The temporal clustering consists of grouping together similar mobile network profiles using the traffic volume generated by BSs at a time slot, this allows to retrieve a reduced number of profiles.…”

“…Under the above mentioned convergence of multiple service and network functions at MEC hosts, and the related high diversity of constraints to take into account to meet the diverse set of requirements, MEC orchestration algorithm scalability and result robustness are key concerns to address. We address the scalability-robustness challenge by extending a problem formulation and related algorithm in [8]. More precisely, we propose the integration of spatial clustering as a precomputation step to the algorithm in [8] to reduce the number of variables of constraints, while integrating in the spatial clustering optimization an objective that aims at making the access point to MEC host assignment more robust against traffic variations within a cluster.…”

“…We address the scalability-robustness challenge by extending a problem formulation and related algorithm in [8]. More precisely, we propose the integration of spatial clustering as a precomputation step to the algorithm in [8] to reduce the number of variables of constraints, while integrating in the spatial clustering optimization an objective that aims at making the access point to MEC host assignment more robust against traffic variations within a cluster. We numerically show for which MEC network sizes the problem becomes tractable.…”

Hyperbolic K-means for traffic-aware clustering in cloud and virtualized RANs

“…Current vRAN systems are split into three parts [26], [27], [29]: the Radio Units (RU) are geographically scattered to achieve coverage; these are connected to Distributed Units (DU), where some layer-1 and layer-2 functions can be performed; the latter are in turn connected to Centralized Units (CU) where base-band operations can be run, along with other RAN orchestration and Mobile-Edge Computing subsystems [4], [13], see Fig. 1c.…”

“…Generally, RU and DU are expected to be collocated or geographically be very close to each other in practice, and transport links provide connections between the DUs and the CU hosting the BBU assigned to the RU-DU subsystem. In 5G and beyond 5G systems, it is envisioned that CU, BBU and MEC servers can be co-located at different edge network sites, typically the so-called Central Office (CO), legacy network points-of-presence being rearchitected nowadays as mini-datacenters [4]. Usually, the distance between an RRH and its BBU-pool (resp.…”