Providing Lifetime Service-Level-Agreements for Cloud Spot Instances

Spot instances are commonly offered by IaaS cloud providers to opportunistically utilize spare capacity and meet temporary user demand for additional resources. Although the availability of service SLAs is a core paradigm of cloud computing, spot instances in practice still come without any service quality guarantees. We aim to extend the spot instance service to provide a probabilistic SLA for eviction probability, based on the user estimate of the maximum expected instance lifetime. This probabilistic foundation simplifies reasoning about the spot instance service and enables providers to construct higher-level SLAs from them. For this to be possible, however, the statistical guarantees must be adhered to strictly, for a wide range of real-world workloads, at cloud scale. We propose a new approach to providing SLAs on the time-until-eviction for spot instances by employing Monte-Carlo simulation to compute the distribution of future spot instance lifetimes at current cloud utilization levels. We then show that an IaaS cloud scheduler can use the quantiles of such conditioned distributions to safely provision spot instance requests and maintain an SLA with a specific target eviction rate.