A performance Brokerage for Heterogeneous Clouds

1.6 Thesis Structure and Summary

This thesis is structured into 8 chapters, including this introduction, and an appendix, organised as follows:

Chapter 2: In chapter 2 we provide definitions of Cloud offerings and deployment models so as to establish a common understanding of them, with particular attention to elasticity and how it can be used to improve performance. We also discuss notions of utility and commodity, which are prevalent in Cloud Computing, and note how performance links both these concepts when applied to Cloud.
Chapter 3: In chapter 3 we discuss some hypothetical examples of how performance variation leads to cost variation in order to provide motivation for the broker. We discuss server hardware and server virtualisation, with particular reference to how performance properties transpire and how providers make use of virtualisation to partition the physical servers into multiple, secure and isolated instances – the primary object of study in this thesis. We review literature related to performance metrics and benchmarks, and note the preference for so-called task progression metrics such as workload execution time or work done in a specified period, over ones which describe machine operations such as number of floating point operations per second. Contributions regarding suitable metrics for performance brokers have been published in: John O’Loughlin and Lee Gillam (2014) "Good Performance Metrics for Cloud Service Brokers". 5th International Conference on Cloud Computing, Grids and Virtualisation (CLOUD COMPUTING 2014).  
Chapter 4: Chapter 4 is dedicated to a review work related to Cloud performance, performance improvement strategies, Cloud brokers, proposals for commodity exchanges for Cloud resources, Cloud pricing including continuous double auctions (CDA), and we discuss how a CDA leads to the equilibrium price. The review will help to ensure validity of structural assumptions regarding brokers and marketplaces. Further, we aim to identify and critique extant performance models, and in doing so determine if they are appropriate for inclusion within the broker model. Contributions in this chapter have been published in John O’Loughlin and Lee Gillam (2015) "Re-Appraising Instance Seeking in Public Clouds" and John O’Loughlin and Lee Gillam (2014) "Performance Evaluation for Cost-Efficient Public Infrastructure Cloud Use".
Chapter 5: The primary purpose of this chapter is to sufficiently characterise performance so as to allow for a realistic model of it to be constructed. To that end we present the results of a number of performance related experiments, the results of which are described in section 1.5.1. Contributions in this chapter have been published in: (1) John O’Loughlin and Lee Gillam (2014) "Should Infrastructure Clouds be Priced Entirely on Performance? An EC2 Case Study" (2) John O’Loughlin and Lee Gillam (2014) "Performance Prediction for Unseen Virtual Machines" (3) Lee Gillam, Bin Li, John O’Loughlin and Anuz Pratap Singh Tomar (2013) "Fair Benchmarking for Cloud Computing Systems".
Chapter 6: This chapter is dedicated to elaborating a model of the broker and determining conditions under which it is profitable. The contributions made in this chapter are described in section 1.5.2 and 1.5.4, some of which have been published in: John O’Loughlin and Lee Gillam (2017) “A Performance Brokerage for Heterogeneous Clouds”.

Chapters 7 and 8 present a discussion and critique of work in this thesis, and present conclusions and opportunities for future work.
Appendix: The appendix contains background material on: (1) CPU architecture and Performance; (2) additional performance results; and (3) consideration of Service Level Agreements (SLAs) and co-location detection. Contributions regarding co-location detection are described in 1.5.3 and have been published in: John O'Loughlin and Lee Gillam (2016) "Sibling Virtual Machine Co-location Confirmation and Avoidance Tactics for Public Infrastructure Clouds".

13.

14.2 Clouds, Utility and Commodity

Cloud Computing covers a large and diverse range of compute, storage and networking services. Due to their scale, ubiquity, and use of standardised offerings the notions of utility and commodity are prevalent in the Cloud literature. However, considering Cloud providers as utilities or Cloud resources as commodities raises a number of questions, such as how to define a common unit of measurement such as the kilowatt-hour, and on what basis different instances should be considered equivalent. We suggest that performance is key to answering both these, for example, instances from different providers may be considered equivalent if they can complete the same amount of computational work per unit of billable time. Similarly, a provider may charge on the basis of amount of work completed. Arguably, notions of Cloud as utility or commodity are intimately related to performance. The objective of this chapter is to provide definitions and characteristics of Cloud services, with particular attention to elasticity and how it can be used to improve performance. We also discuss the extent to which Cloud can be considered as either utility or commodity.

In sections 2.1 and 2.2 we begin by defining and discussing the basic characteristics and properties of Cloud services. In sections 2.3 and 2.4 we discuss the various service and deployment models, which define what the service is and where the service is offered from. These sections provide a common understanding and terminology for Cloud service and deployment models.
In section 2.5 we describe how a typical Cloud is comprised of multiple sub-Clouds (Regions) which are located in different geographical locations which are very loosely coupled and so can be considered essentially independent of each other. As Regions come on-line at different times, and grow and expand, they are not necessarily comprised of the same types of hardware as each other and yet they purportedly offer the same types of resources for rent. Looking ahead to chapter 5 we will see how these differences lead to performance differences across Regions.
In section 2.6 we discuss how providers make instances available in a variety of non-negotiable flavours and sizes known as instances types. Standardisation likely eases the mass production of instances, and draws natural comparisons with commodity, and we note that offerings from different providers are increasingly similar, indicative of a market undergoing commoditisation. In section 2.7 we discuss utility and commodity, taking the opportunity to provide some background with regards to previous generations of distributed systems.
We begin with a review of how Cloud services are defined.