An Algorithm for Handling Starvation and Resource Rejection in Public Clouds

Abstract

Nowadays, Cloud Computing is emerging as a replacement for traditional physical hardware computing. Infrastructure-as-a-Service (IaaS) is one of the fundamental cloud computing models, where users can request virtual resources with various capabilities whenever needed. Haizea is a well-known open source virtual machine scheduler for IaaS clouds, offering four types of leases, i.e. requests for virtual machines: immediate, Best-effort (BE), Dead Line Sensitive (DLS) and Advanced Reservations (AR). One of Haizea disadvantages is that BE leases will be preempted whenever their resources are required by AR or Immediate leases. Thus, when the system has a high number of AR or immediate lease requests, BE leases will wait for a long time, or even forever, resulting in what is known as the starvation problem. To prevent this problem, some policies suggest rejecting AR lease requests if BE leases have been waiting for a long time. However, this would result in decreased resource utilization and customer satisfaction. Therefore, this paper proposes an enhancement to Haizea scheduler that reduces the starvation problem and request rejection rate, while increasing resource utilization of the cloud infrastructure. Simulation results demonstrated that the proposed algorithm has outperformed pure Haizea in all of the aforementioned performance measures.