An Online Algorithm for Cost Minimization of Amazon EC2 Burstable Resources

Many computational workloads do not require a constantly high CPU. They generally have occasional peak computation needs with low CPU utilization most of the time. Cloud service providers have started providing burstable instances to cater to such workloads. Burstable instances allow users occasional surges in utilization provided the job’s CPU utilization remains below a certain baseline usage most of the time. Instead of running a job by provisioning regular instances based on peak usage, using burstable instances can significantly reduce costs. Amazon EC2 provides t-instances (T2/T3/T4g etc.) with different baseline usage (like 5%, 10%, 20%, 30%) to the users with two modes – (standard and unlimited). Burstable instances have a credit mechanism with a maximum capacity limit. When the usage is below the baseline, a credit is earned that can be utilized later during peak demands. In the present work, we attempt to utilize different types of burstable instances for workloads with a provision to migrate from one baseline to another to optimize (i) (minimize) surplus charges and (ii) (maximize) the use of credit earned. We have implemented the proposed migration approach on AWS burstable instances with real PlanetLab demand traces. We show that even without prior knowledge of future CPU utilizations of a job, our proposed online algorithm can achieve cost comparable to a hypothetical future-aware algorithm that assumes future knowledge of utilization values.