Return to search

Provenance-driven diagnostic framework for task evictions mitigating strategy in cloud computing

Cloud computing is an evolving paradigm. It delivers virtualized, scalable and elastic resources (e.g. CPU, memory) over a network (e.g. Internet) from data centres to users (e.g. individuals, enterprises, governments). Applications, platforms, and infrastructures are Cloud services that users can access. Clouds enable users to run highly complex operations to satisfy computation needs through resource virtualization. Virtualization is a method to run a number of virtual machines (VM) on a single physical server. However, VMs are not a necessity in the Clouds. Cloud providers tend to overcommit resources, aiming to leverage unused capacity and maximize profits. This over-commitment of resources can lead to an overload of the actual physical machine, which lowers the performance or lead to the failure of tasks due to lack of resources, i.e. CPU or RAM, and consequently lead to SLA violations. There are a number of different strategies to mitigate the overload, one of which is VM task eviction. The ambition of this research is to adapt a provenance model, PROV, to help understand the historical usage of a Cloud system and the components contributed to the overload, so that the causes for task eviction can be identified for future prevention. A novel provenance-driven diagnostic framework is proposed. By studying Google’s 29-day Cloud dataset, the PROV model was extended to PROV-TE that underpinned a number of diagnostic algorithms for identifying evicted tasks due to specific causes. The framework was implemented and tested against the Google dataset. To further evaluate the framework, a simulation tool, SEED, was used to replicate task eviction behaviour with the specifications of Google Cloud and Amazon EC2. The framework, specifically the diagnostic algorithms, was then applied to audit the causes and to identify the relevant evicted tasks. The results were then analysed using precision and recall measures. The average precision and recall of the diagnostic algorithms are 83% and 90%, respectively.
Date January 2017
CreatorsAlbatli, Abdulaziz Mohammed N.
ContributorsLau, Lydia ; Xu, Jie
PublisherUniversity of Leeds
Source SetsEthos UK
Detected LanguageEnglish
TypeElectronic Thesis or Dissertation

Page generated in 0.2705 seconds