MLF-DRS: A Multi-level Fair Resource Allocation Algorithm in Heterogeneous Cloud Computing Systems

Abstract

Cloud computing is a novel paradigm which provides on demand, scalable and pay-as-you-use computing resources in a virtualized form. With cloud computing, users are able to access large pools of resources anywhere without any limitation. In order to use the provided facilities by the cloud in an efficient way, the management of resources is an undeniable fact that should be considered in different aspects. Among all those aspects, resource allocation has received much attentions. Given the fact that the cloud is heterogeneous, the allocation of resources has to become more sophisticated. As a first promising work to deal with that problem, Dominant Resource Fairness (DRF) has been proposed which takes into account dominant shares of users. Although DRF has a sort of desirable fairness properties, it has some limitations that have already been identified in the literature. Unfortunately, DRF and its recent developments are not intuitively fair with respect to various resource demands. In this paper, we propose a Multi-level Fair Dominant Resource Scheduling (MLF-DRS) algorithm as a new allocation model inspired by Max-Min fairness and proportionality. Unlike other works that they equalize dominant shares of different resource types which leads to starvation in the maximization of allocation for some users, our algorithm guarantees that each user receives the resources they desire for based on dominant shares. As can be deducted from the mathematical proofs, MLF-DRS provides a full utilization of resources and meets some of the desirable fair allocation properties and it is applicable to be used in a naïve extension form in the presence of multiple servers as well.

H-FFMRA: A Multi Resource Fully Fair Resources Allocation Algorithm in Heterogeneous Cloud Computing

“…If U k d s , and U k ds denote a group of users with specific dominant, and nondominant resources in each server sequentially. Then, the proportion of ϕ d , and ϕ d are used to calculate final shares, employing MLF-DRS [14] based on the following optimization problem.…”

“…The formulations (12), and ( 13) indicate that a group of users in each server receive an equal proportion of the resource pool. To measure resource allocation with fairness with regards to each user in each server, the Jain's index is applied for each server based on (14), and (15) for dominant and non-dominant shares.…”

Minimizing Resource Waste in Heterogeneous Resource Allocation for Data Stream Processing on Clouds

“…In the allocating phase, DRFH allocates a similar proportion of the dominant resource share ratio among application demands and available resources on each server to reduce the amount of unused resources. Hamzeh et al [41] proposed a Multi-level Fair Dominant Resource Scheduling (MLF-DRS) algorithm to guarantee the fairness of resource demands based on dominant shares. However, previous approaches were too complicated and are not proposed for the data streaming application in Spark which mainly focuses on the in-memory processing.…”

A New Approach to Calculate Resource Limits with Fairness in Kubernetes

“…Finally, according to (12) and (13), the allocated resource to each user is calculated as the final stage of MLF-DRS.…”

“…Taking into account the above-mentioned problems, In this paper, the first attempt is to model and integrate three different fair allocation algorithms, MLF-DRS [13] and FFMRA [14] as well as DRF in Kubernetes, trying to assign resource limits fairly among different pods running in a specific node.…”