Network-Aware Resource Allocation for Cloud Elastic Applications

  • Fatima Mohammed Alqayedi

Student thesis: Doctoral Thesis

Abstract

Cloud elastic applications and services are widely deployed these days. These elastic applications have the ability to scale up and down the computing resources according to the fluctuation of the workload. Cloud computing resources for these applications should be allocated by the cloud provider in an efficient way so that minimal resources are allocated to satisfy SLO (Service Level Objective) requirements. This thesis proposes algorithms and techniques to efficiently allocate resources for cloud-hosted applications, and at the same time be able to meet the SLA (Service Level Agreement) objectives, taking into account network utilization being used by other applications and services hosted in the cloud. Such network utilization and activities can significantly impact the SLO response time experienced by the cloud users. Our proposed algorithms reduce the cost for both clients and service providers. Clients pay for the usable resources and the service providers provide minimum needed resources. More specifically, we propose an Adaptive Cloud Resource Allocation scheme (ACRAS) that allocates minimal cloud Virtual Machine (VM) resources that are needed to satisfy a given SLO response time for cloud-based elastic applications. More importantly, the algorithm attempts to mitigate any response time violation that could arise during the provisioning of cloud VM instances. Our proposed scheme utilizes queueing theory to estimate the number of VM instances that are needed to satisfy the response time according to the current workload. The scheme also employs reactive provisioning technique to examine workload conditions, and re-compute needed VM instances at a periodic interval at CPU utilization, and workload of allocated VMs are used as a threshold to adjust (to add or remove) the number of allocated VMs. Furthermore, we propose a heuristic network-aware virtual machine placement technique. The technique considers end-to-end delay for VMs and delay of existing network traffic of other applications running on the same cloud infrastructure. Two approaches are proposed which are classic first fit approach and network-aware approach and compared on response time violation and time execution. The classical first fit approach addresses the classic first fit algorithm that focuses only on addressing compute resource requirements. On the other hand, the network-aware approach incorporates network awareness to the first fit algorithm. Prior to the actual placement of VMs, the algorithm computes the end-to-end delay to ensure that the traffic caused by the VM does not affect other running applications in cloud data center. We have cross-validated our analytical models and algorithms using simulation, with real workload traces obtained from World Cup 98. The results show that our scheme is very effective in satisfying SLO response time while minimizing violations during spikes of workload or during provisioning of VMs.
Date of AwardJul 2017
Original languageAmerican English

Keywords

  • Cloud computing
  • elasticity
  • network-aware
  • elastic applications
  • resource allocation
  • SLO response time
  • virtual machine placement
  • response time violation.

Cite this

'