Content area
Burstable instances provide a low-cost option for consumers using the public cloud, but they come with significant resource limitations. They can be viewed as "fractional instances" where one receives a fraction of the compute and memory capacity at a fraction of the cost of regular instances. The fractional compute is achieved via rate limiting, where a unique characteristic of the rate limiting is that it allows for the CPU to burst to 100% utilization for limited periods of time. Prior research has shown how this ability to burst can be used to serve specific roles such as a cache backup and handling flash crowds.
Reserved burstable instances offer additional cost savings over burstable instances by requiring a commitment of either one or three years. Their cost is divided into an upfront payment and an hourly discounted rate for the duration of the term. It is possible to reserve instances with no upfront payment but it offers the lowest amount of discount. Full upfront payment offers the highest amount of discount, up to 72% compared to regular on-demand instances. Prior works have shown how reserved instances can be used with regular on-demand instances to optimize cost but no research has been conducted for reserved burstable instances. This is fundamentally different from reserved instances because of the burst capacity of burstable instances has to be taken into account for reserved burstable instances.
This dissertation investigates how to generally utilize burstable and reserved burstable instances to optimize cost and meet latency SLOs via their burst capability. In the first part of the dissertation, we demonstrate how the system AutoBurst is able to utilize burstable instances to achieve this by controlling both the number of burstable and regular instances along with how they are used. Evaluations in AWS show that AutoBurst is able to reduce cost by up to 25% while maintaining latency SLOs. In the second part of the dissertation, I show how reserved burstable instances can be utilized along with reserved and regular on-demand instances for additional cost savings for long-term workloads. I propose a Linear Programming formulation using reserved burstable instances to solve for a near-optimal solution in the long-term when the demand is known. I also propose a heuristic solution ResBurst that is able to outperform reserved only solutions even when the future demand is mispredicted. Evaluations in AWS show that using reserved burstable instances saves cost over using only reserved instances.