Resource allocation for edge computing with multiple tenant configurations

Abstract

Edge Computing (EC) consists in deploying computational resources, e.g., memory, CPUs, at the Edge of the network, e.g., base stations, access points, and run there a part of the computation currently running on the Cloud. This approach promises to reduce latency, inter-domain traffic and enhance user experience. Since resources at the Edge are scarce, resource allocation is crucial for EC. While most of the studies assume users interact directly with the Edge submitting a sequence of tasks, we instead consider that users will interact with different Service Providers (SPs), as they currently do in the Web. We therefore consider the case of a Network Operator (NO) that owns the resources at the Edge and must decide how much resource to allocate to the different tenants (SPs). We propose MORA, a polynomial time strategy which allows the NO to maximize its utility, which can be inter-domain traffic savings, improved users' QoE or other metrics of interest. The core of MORA is that (i) it exploits service elasticity, i.e., the fact that services can adapt to the resources allocated by the NO and rely on a remote Cloud for the excess of computation, (ii) it is suitable for micro-services architecture, which decomposes a single service in a set of components, which MORA places in the different computational nodes of the Edge and (iii) it copes with multi-dimensional resources, e.g., memory and CPUs. After analyzing the properties of the algorithm, we show numerically that it performs close to the optimum. To guarantee reproducibility, the numerical evaluation is performed on publicly available traces from Google and Alibaba clusters and in synthetic scenarios and our code is open source. CCS CONCEPTS • Networks → Cloud computing; Network management; Programmable networks; • Computer systems organization → Cloud computing; n-tier architectures;

Container Placement and Migration in Edge Computing: Concept and Scheduling Models

“…However, this increases the container file size. Containerizing microservices has been found to be better than running multiple microservices in one container [10]. Although this increases the complexity of orchestration, it boosts the container boot up time because all the microservices can be booted simultaneously.…”

“…Computing can equally be migrated to the cloud from an edge server, in a situation where there is no suitable edge node to transfer services to, and then back to an edge node when one becomes available [67]. Where an edge node cannot complete a computation due to limited resources, part of that computation can be moved to the cloud [10]. Containers are used in cloud servers and so seamless migration is possible between edge node and cloud server.…”

“…In the container placement problem in edge computing, the objective function relates to the goals such as minimizing latency [84], minimizing overall cost of providing edge services [12,39], minimizing total energy used by the server [86] and minimizing load imbalance [39]. Satisfying SLA is a maximization objective [10,33]. Many of the problems tend towards a joint optimization problem, see examples in [14,39,71,[84][85][86].…”

“…The decision variables are the scheduling decisions to be taken, such as which computing tasks to accept [10,87], what containers will the task be assigned to [87], what edge nodes will the containers be deployed [33] or migrated [39]. Constraints in optimization problems can be described as finite resources that the decision maker has to consider when taking the decision.…”

“…From a service provider standpoint, the scheduler decides how the different incoming computing requests are mapped to containers, and how containers are mapped to heterogenous edge nodes, to service the request. These mappings attempt to fulfil a set of objectives within given operational constraints [10][11][12][13]. The scheduling decision is a complex one.…”