Workload-Aware Provisioning in Public Clouds

Public cloud services rely on virtualization to support multitenancy--that is, allowing customers from different organizations to share the datacenter infrastructure. Unfortunately, today's public clouds fail to provide sufficient isolation. Hardware resources are often multiplexed between virtual machines belonging to different customers, which can cause performance interference. This article characterizes interference on the network latency between virtual machines, and shows that the Amazon Elastic Compute Cloud (EC2) suffers from a long-tail latency problem due to coscheduling of CPU-bound and latency-sensitive tasks. The Bobtail system allows cloud customers to proactively detect and avoid these bad neighboring virtual machines without help from cloud service providers.