Fairness for resource allocation in cloud computing.

Abstract

Resource allocation with fairness, considering multiple types of resources has become a substantial concern in state-of-the-art computing systems. Accordingly, the rapid growth of cloud computing has highlighted the importance of resource management as a complicated and NP-hard problem. Unlike traditional frameworks, in modern data centers, incoming jobs pose demand profiles, including diverse sets of resources such as CPU, memory, and bandwidth over multiple servers. Therefore, the fair distribution of resources respecting such heterogeneity has appeared to be a challenging issue. In this thesis, the major research gaps are identified, associated with recent solutions in resource allocation and scheduling with fairness in cloud computing. The recent developments have satisfied some desirable fairness properties such as sharing-incentive, Pareto- efficiency, envy-free, and strategy-proof in general perspective. However, some of these promising features have not been satisfied for some users with demands dominated on a particular resource type that cause an allocation to be intuitively unfair. First, recent approaches have ignored the boundaries of fair-share and resource demands in allocation decisions. Additionally, those approaches have fallen short in prioritizing tasks, considering different resource types in scheduling time that may result in increasing the response time for some users. Second, the recent developments have considered the equalization only for dominant resources that could be an obstacle against utility maximization for a certain number of users. Besides, there is no specific measure for evaluating the fair distribution of resources among tasks with dominant and non-dominant resources, in single and heterogeneous server settings. Third, it is still unclear how the number of dominant resources in multiple servers, considering a specific resource type may affect the Pareto-efficiency and sharing incentive properties. Fourth, investigating the fairness problem, taking into account multiple resource types seems like a missing point in the Kubernetes environment. This significant issue may increase the response time due to a considerable number of pod evictions. This thesis seeks to address these substantial gaps. First, a new mechanism is introduced to calculate shares among users, considering a fair-share function that solves a utility maximization problem. Furthermore, a new queuing mechanism is proposed to reduce response time for incoming tasks dominated on different resource types. Second, to address the fair distribution of resources among users, a fully-fair allocation algorithm is presented as well as a new fairness measure for multi-resource environments. Third, to tackle issues concerning Pareto-efficiency and sharing-incentive in heterogeneous servers, a novel task scheduling mechanism is introduced. Fourth, a new model is suggested to tackle the fairness problem in the Kubernetes framework to reduce the number of pod evictions in scheduling time. The experiments conducted in the Cloudsim framework, using randomly generated work- loads and Google workload traces show that our proposed algorithms achieve higher utilization of resources as well as fairness compared to DRF. Moreover, the evaluations indicate that the proposed algorithms satisfy desirable fairness properties.