Using Workload Prediction and Federation to Increase Cloud Utilization

The wide-spread adoption of cloud computing has changed how large-scale computing infrastructure is built and managed. Infrastructure-as-a-Service (IaaS) clouds consolidate different separate workloads onto a shared platform and provide a consistent quality of service by overprovisioning capacity. This additional capacity, however, remains idle for extended periods of time and represents a drag on system efficiency.The smaller scale of private IaaS clouds compared to public clouds exacerbates overprovisioning inefficiencies as opportunities for workload consolidation in private clouds are limited. Federation and cycle harvesting capabilities from computational grids help to improve efficiency, but to date have seen only limited adoption in the cloud due to a fundamental mismatch between the usage models of grids and clouds. Computational grids provide high throughput of queued batch jobs on a best-effort basis and enforce user priorities through dynamic job preemption, while IaaS clouds provide immediate feedback to user requests and make ahead-of-time guarantees about resource availability.We present a novel method to enable workload federation across IaaS clouds that overcomes this mismatch between grid and cloud usage models and improves system efficiency while also offering availability guarantees. We develop a new method for faster-than-realtime simulation of IaaS clouds to make predictions about system utilization and leverage this method to estimate the future availability of preemptible resources in the cloud. We then use these estimates to perform careful admission control and provide ahead-of-time bounds on the preemption probability of federated jobs executing on preemptible resources. Finally, we build an end-to-end prototype that addresses practical issues of workload federation and evaluate the prototype's efficacy using real-world traces from big data and compute-intensive production workloads