Some Studies on Renewable Energy-Based Scheduling Algorithms for Geo-Distributed Datacenters

Abstract

The cloud marketplace is continuously rising as enterprises desire to streamline their processes. The marketplace can address the potential computation, storage, databases, and bandwidth needs across various sectors, including business, industry, manufacturing, en tertainment, government, education, agriculture, banking and smart cities. As adaptability increases, cloud service providers (CSPs) expand their datacenters to handle requests of any size. It increases the fossil fuels consumed in each datacenter, increasing the over all cost. Therefore, many CSPs have adopted renewable energy (RE) sources to increase profitability and reduce carbon emissions. However, RE generation fluctuates with time, lo cation and climate conditions, which introduces uncertainty (UN) in fulfilling user requests (URs). Thus, non-renewable energy (NRE) generation continues to power the datacenters to make stability. Recent works are directed to the use of RE generation followed by NRE generation while assigning the URs to the resources of the datacenters. However, they present the requirements of URs using the processor nodes without considering memory nodes. Moreover, these works aim to maximize the usage of RE or minimize the cost and do not model the UN of RE and NRE resources and the level of UN (UNL). Furthermore, the URsmayberedirected from one datacenter to another depending on the presence of RE resources to minimize the cost. However, systematically planning of datacenter migration to move URs is quite challenging. This thesis mainly addresses the UR-based scheduling problems in geo-distributed data centers. It presents several algorithms for these problems, considering processor and mem ory nodes, UN, UNL, and migration. Firstly, we present two algorithms, processor and memory-based future-aware best fit (PM-FABEF) and processor and memory-based high est available renewable first (PM-HAREF), that incorporate both processor and memory nodes for geographical load balancing (GLB). PM-FABEF determines the cost of proces sor and memory nodes for assigning URs to the datacenters and assigns them to the least cost datacenter. PM-HAREF determines the highest RE resource slots in processor and memory for assigning the URs. Secondly, we present three UR-based scheduling algo rithms, namely UN-based future-aware best fit (UN-FABEF), UN-based highest available iii renewable first (UN-HAREF) and UN-based round-robin (UN-RR), by managing the UN of RE and NRE. These algorithms consider two types of UN, namely UN of RE (UN-RE) and UN of NRE (UN-NRE) resources, concerning the UR. UN-FABEF matches the UR to all the available datacenters and determines the assignment cost, including UN time du ration. Then, it assigns that UR to a datacenter that results in the least assignment cost. UN-HAREFmatches the UR to all the available datacenters and determines the number of RE resource slots, including UN time duration. Then, it assigns that UR to a datacenter that results in the highest number of RE resource slots. On the contrary, UN-RR assigns the URs to the datacenters in a circular fashion. Thirdly, we extend the three benchmark algorithms, namely FABEF, HAREF and RR, by incorporating UN and UNL, and we call them UNL-based future-aware best fit (UNL-FABEF), UNL-based highest available re newable first (UNL-HAREF) and UNL-based round-robin (UNL-RR), respectively. The goal of UNL-FABEF is to minimize the overall cost, whereas UNL-HAREF is to maxi mize the available RE usage. On the contrary, UNL-RR assigns the URs to the datacenters in a roundabout fashion. Then, we introduce the UNL-based multi-objective scheduling algorithm (UNL-MOSA) to make a trade-off between UNL-FABEF and UNL-HAREF. UNL-MOSAcreates abalance between the overall cost and the available RE usage. Lastly, weintroduce a novel RE-oriented migration algorithm (REOMA) to minimize the total cost of geo-distributed datacenters through strategic migration between datacenters. REOMA f inds the cost based on the time window of the UR in each datacenter. Then, it fits the cost of each datacenter into a polynomial curve based on the time window and determines the slope and intercept. Subsequently, it finds the migration points between the datacenter with the lowest cost and the other datacenters and performs the migration between a pair of datacenters that results in the lowest cost. We perform rigorous simulations on the proposed algorithms and measure their per formance in terms of various performance metrics, namely overall cost (OCO), the total number of used RE resource slots (TNRE), the total number of used NRE resource slots (TNNRE), the UN time (UNT) and the UN cost (UNCO). The proposed algorithms are compared with three benchmark algorithms using fifty instances of ten datasets with 200 to 2000 URsand20to200datacenters based on their applicability. For comparison purposes, iv we also evaluate the simulation results against existing scheduling algorithms and present the findings in various tabular and graphical formats. The comparison results demonstrate that the proposed scheduling algorithms outperform existing ones in terms of the afore mentioned performance metrics. Furthermore, we validate these results using analysis of variance (ANOVA) statistical tests based on the applicability. Keywords: Cloud Computing, Cloud Service Provider, Datacenter, Geo-Distributed Datacenters, Geographical Load Balancing, Non-Renewable Energy, Overall Cost, Renew able Energy, Scheduling, Uncertainty, Uncertainty Level, User Request