טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
Ph.D Thesis
Ph.D StudentAgmon Ben-Yehuda Orna
SubjectEfficient, Non-Cooperative Sharing of Computing Resources
DepartmentDepartment of Computer Science
Supervisor Professor Assaf Schuster
Full Thesis textFull thesis text - English Version


Abstract

Shared computing resources such as clouds and grids are liable to be inefficiently utilized due to conflicts of interest: the hardware and electricity bill are paid for by one economic entity, while the workloads that make use of the resources benefit other economic entities. We begin by examining resource utilization strategies available to clients in a traditional environment of grids, assisted by an on-demand Infrastructure-as-a-Service (IaaS) cloud. The less reliable the grids are, the more wasteful the strategies are, because they involve task replication. Another unreliable environment that requires strategic client behavior is Amazon's EC2 cloud spot instances. The spot instances' unreliability is the result of their changing prices. These prices are supposed to reflect changes in supply and demand, but we discovered that most of the perceived unreliability was artificially generated, masking resource under-utilization. Amazon was, for a while, the spearhead of several trends in the public cloud industry. These trends include refining the resource rental-time granularity, refining the resource quantity granularity, and offering more flexible service level agreements. Other providers soon followed suit, pushing these trends further by reducing the rental-time granularity to minutes and the resource granularity to hundreds of MBs and CPU fractions. Due to the need for non-cooperative users to efficiently share resources, these trends will likely culminate in the rise of a new economic model that we term the Resource-as-a-Service (RaaS) cloud. Finally, we propose a RaaS prototype that efficiently rents physical memory to non-cooperative virtual machines at a fine time and resource granularity, at personally adapted prices, and with flexible service level agreements.