Department of Computer Engineering
GRADUATE RESEARCH SEMINAR
Cloud Job Stacked Scheduling for Reduced Cost and Improved Time
Computer Engineering Department
Virtual machine allocation is an open ended problem. For the cloud provider, it can be optimized by having extra knowledge, such as network communication pattern, resource usage, and total budget and so on. However, from the user's perspective, actual VM allocation in datacenters cannot be optimized efficiently, because the VM placement is not at the hands of the user. In public clouds, users only specify their needs, and it results in allocated VMs. Since each VM has an hourly cost, allocation must be done carefully. If we have a small task that finishes in less than hour on two machines it is illogical and not required to choose the expensive one. In this paper, we show that in certain workloads, assigning multiple jobs to a single VM can reduce the cost or improve time or both, even if a VM is overly utilized. We show that given an utilization model, it is beneficial to schedule some tasks together on VM rather than individual scheduling.
Additionally, we support this model with real experiments on our test bed.
DATE: 16 March, 2015, Monday @ 15:40