Smart spot instances for the supercloud

Abstract

In this paper, we explore the use of live VM migration to take advantage of spot markets such as provided by Amazon and Google. These markets provide an exciting low cost alternative to regular VM instances, but the threats of price spikes and premature termination severely limit their usability. Migration can address these threats: spot market instances facing price hikes or termination can migrate to other instance types, including regular ones. Reliability can be further improved by replication. In this paper we investigate various design options and present some preliminary results of experiments with dynamic programming techniques, both using simulation and using a realistic deployment. We find that in unstable markets we can achieve significant savings at low overhead and while maintaining good reliability. 1 Our Amazon account does not grant access to the us-east-1a zone.

Provisioning Spot Instances Without Employing Fault-Tolerance Mechanisms

“…Most of the prior works focused on reducing the effect of spot instance revocations using fault-tolerance methods, such as replication [9]- [11], checkpointing [4], [6], [12], and VM migration [8], [13]. Voorsluys et al [9] proposed a fault-aware resource allocation approach that applies the price of spot instances, runtime estimation of applications, and task duplication mechanisms to economically run batch jobs in spot instances.…”

Economics of a supercloud

“…Instead, the Supercloud enables migrating to different types of virtual machine instance (e.g. from 2XL to 4XL instances) (Jia et al 2016).…”

SPA: Harnessing Availability in the AWS Spot Market

“…Although transient instances do no offer uptime guarantees, many studies in the past have explored the benefit of these instances. For example, SuperCloud [9] explores how to utilize transient resources via algorithms that deal with the expected failures by migrating the instance running state from one virtual machine to another. Other approaches include saving the service data in a shared partition [8], [10], [16], checkpointing the job processing at regular time intervals [23], increasing application availability by proactively migrating applications between spot and on-demand instances [17]- [19] and mix-and-match spot and on-demand instances to trade availability for cost reduction [2].…”