Resource-Aware and Load-Balanced Scheduling for Cloud Computing Environments


High Performance Computing (HPC) is becoming ever more signi cant research tool in modelling complex scienti c problems. Being large-scale nature, scheduling techniques are key to e ectuate the execution in Cloud environment at the lowest possible cost without the need of owing any physical infrastructure. Meanwhile, the provision of high-performance computing by employing fewer resources with a cap on nancial constraints is also emerging as an appealing research area in Cloud. While constructing Cloud datacenters, we should not only adopt to the best suitable software and hardware practices, but also have to keep a close eye on datacenter operations, energy consumption, scalability, security, technological innovation, fault tolerance, and many other issues. In addition, the resources should be utilized in an ecient manner along with considering the wisdom of general framework of Cloud scheduling.

In this thesis, a computation-aware load balancing scheme known as Resource-Aware Load Balancing Algorithm (RALBA) is proposed for scheduling computeintensive, independent, and non-preemptive jobs on a compute Cloud. Using RALBA, the HPC workload can be distributed in a load-balanced manner guaranteeing improved resource utilization. RALBA scheme is designed with the intention to contemplate the total computing capabilities (i.e., computing powers) of Virtual Machines (VMs) and the computing requirements of the submitted HPC workload, which provides a load-balanced schedule with improved utilization of virtual computing resources and ultimately that of underlying physical machines in Cloud datacenters. RALBA scheme is evaluated on varying workload compositions (i.e., using several synthetic and benchmark scienti c datasets) by employing di erent computing environments (i.e., heterogenous computing machines). The performance outcomes of RALBA have revealed that it provides substantial improvement against traditional and state-of-the-art scheduling algorithms in terms of makespan, throughput, and resource utilization. Moreover, the potential of eight state-of-the-art algorithms against RALBA is investigated in terms of makespan and resource utilization in amalgamation with machine-level load balancing using nine di erent instances of two benchmark scienti c datasets (i.e., employing jobs heterogeneity) on seven di erent computing architectures The empirical examination endorsed that RALBA produces a better load-balanced schedule ensuring to release all the computing resources within almost similar quantum of time, which e ectuate the Cloud Service Providers (CSPs) to reduce the execution delays effecting the next batch of Cloud jobs.

In addition, the HPC datasets are increasingly becoming more pertinent when executing resource allocation, and load balancing techniques for an eagle-eyed examination of ecacy and performance on Cloud. However, a real Cloud dataset is hard to acquire for such purpose due to users’ data con dentiality and policies maintained in SLAs by Cloud service providers. Therefore, a new dataset known as Google Cloud Jobs (GoCJ) is also proposed in this dissertation as an alternative to benchmark workloads for scheduling and resource provisioning in a compute Cloud. The GoCJ dataset is realistic dataset generated using Monte Carlo simulation considering the real workload as perceived in Google cluster traces.

SLA standardizes the assumptions and responsibilities of both the Cloud users and CSPs that acts as a roadmap to the successful implementation of Cloud services. Therefore, an extension of RALBA is also presented in the form of SLA and Resource-Aware Load Balancing Algorithm (SLA-RALBA) for heterogeneous Cloud, which is a cost-ecient and computation-aware load balancing technique. The empirical results evidently revealed that SLA-RALBA provides improved balance between execution time and cost by guaranteeing a drastic improvement in resource utilization on Cloud as compared to existing cost-ecient SLA-techniques.

Download full paper