MNO與MVNO於雲端-邊緣運算之優化資源分配

Abstract

在Mobile Network中，大型的Mobile Network Operator (MNO)擁有大量的運算資源供消費者租用，但MNO本身能接觸到的客群較為有限，因此搭配Mobile Virtual Network Operator (MVNO)的合作是一個常見的商業營運模式。MNO將部分的運算資源租給本身不設置設施的虛擬網路服務業者(MVNO)，使其也能夠提供網路服務，並協助其營運以擴大自身市場客群。在此營運模式下，MVNO本身不建置任何基礎建設(如基地台、運算主機)，其服務資源均是向MNO租用，故此議題涉及MNO與MVNO的資源分配效率與商業模式之合理性，是而成為新的研究議題。 本論文要解決的問題主要有兩項，第一個是運算資源的分配問題，在MNO決定與MVNO合作時，雙方必須簽訂合約以決定在合約時間內MVNO要租用一定數量的運算資源，即MNO要在合約範圍內決定分配多少Virtual Machine (VM)給MVNO，增加自身營收並同時讓MVNO合理營運，我們稱這個問題為VM Allocation。在此階段本論文假設MNO可以透過蒐集過去的營運資料來進行估算，以進行最符合雙方商業利益的VM分配。 在合約簽訂後每回合的營運中，除前述VM分配外，MNO及MVNO會根據其各自客戶需求提供運算服務，而本論文要解決的第二個問題是工作分配問題，每當一個運算工作進入系統，會立即對該工作進行分配，決定該工作要放到哪一台VM進行運算，並且考慮cloud VM與edge VM之間的差異性，我們稱這個問題為Task Assignment。另外我們將運算工作分為VoIP, IP Video與FTP三種類別，並設定每個VM只能服務同一種類別的工作，每個工作類別有其獨特的流量特性，根據其特性的不同制定不同的效用函數，task assignment的目標為在使用者被分配到的資源量至少要大於資源下限(operating value)的前提下，最大化使用者對於該工作所獲服務的滿意度(utility)，根據operator營運策略的不同，他們制定的operating value也會不同，代表其不同的商業考量。 本論文參考Google在2019年釋出的真實資料ClusterData trace，摘取其參數特性並在套用本論文所指定的隨機程序來進行模擬，以貼近真實世界的運作。因此本論文的主要貢獻為設計貼近現實的環境與模型，探討如何公允且自動的分配資源給MNO與MVNO，並達到雙邊都可接受的最佳營運效果。

In Mobile Network, Mobile Network Operator (MNO) is a big company that owns large amount of computing resource for customers to rent, but MNO itself can only reach a limited range of customers, so cooperating with another small company such as Mobile Virtual Network Operator (MVNO) is a common business mode is real world. Under this business mode, MNO rent part of its computing resource to MVNO that do not have their own facilities, so that MVNO can provide computing services to their own customers. This is a new research topic that involves the efficiency of resource allocation between MNO and MVNO and the rationality of the business mode. There are two main problems to be solved in this thesis. The first one is the allocation of computing resources. When MNO decides to cooperate with MVNO, both parties must sign a contract to decide that MVNO will rent a certain amount of computing resources, that is, MNO needs to decide how many Virtual Machines (VMs) to allocate to MVNO, to increase their own revenue and allow MVNO to operate reasonably. We call this problem as VM Allocation. We assume that MNO can make estimates by collecting past operational data to make the VM Allocation mechanism allocate the fittest amount of resources for both parties. After the VM Allocation, MNO and MVNO will provide computing services to their own customers, the second problem to be solved in this paper is tasks assignment. Whenever a task enters the system, we will assign the task immediately, decide which VM the task should run at, we also consider the difference between the cloud VM and the edge VM. We call this problem as Task Assignment. We divide computing application tasks into three categories: VoIP, IP Video and FTP, and assume each VM can only serve tasks of the same category. Each task category has unique traffic characteristics, and different utility functions are designed accordingly. The goal of Task Assignment is to maximize the satisfaction (utility) of users with the resource provided by a VM under the premise that the amount of resources allocated to the user is enough but not too many. This thesis refers to the real trace released by Google in 2019 called ClusterData. We extract its parameter characteristics and design a random program for simulating. The main contribution of this paper is to design an environment and model that close to realistic, discuss how to allocate resources between MNO and MVNO fairly and automatically, and achieve the best operating effect that is acceptable to both parties.