Portfolio-driven Resource Management for Transient Cloud Servers

Abstract

Cloud providers have begun to offer their surplus capacity in the form of low-cost transient servers, which can be revoked unilaterally at any time. While the low cost of transient servers makes them attractive for a wide range of applications, such as data processing and scientific computing, failures due to server revocation can severely degrade application performance. Since different transient server types offer different cost and availability tradeoffs, we present the notion of server portfolios that is based on financial portfolio modeling. Server portfolios enable construction of an "optimal" mix of severs to meet an application's sensitivity to cost and revocation risk. We implement model-driven portfolios in a system called ExoSphere, and show how diverse applications can use portfolios and application-specific policies to gracefully handle transient servers. We show that ExoSphere enables widely-used parallel applications such as Spark, MPI, and BOINC to be made transiency-aware with modest effort. Our experiments show that allowing the applications to use suitable transiency-aware policies, ExoSphere is able to achieve 80% cost savings when compared to on-demand servers and greatly reduces revocation risk compared to existing approaches. ACM Reference format:Prateek Sharma, David Irwin, and Prashant Shenoy. 2017. Portfolio-driven Resource Management for Transient Cloud Servers.

Cloud-scale VM-deflation for Running Interactive Applications On Transient Servers

“…Systems for handling transient server revocation use a combination of fault tolerance and resource allocation to mitigate the performance and cost effects of preemptions. Prior work has focused on system [43,45] and application [23,30,40,41,49,51] support for handling preemptions. We believe that deflatable VMs minimize the need for such middleware, and can avoid the performance, development, and deployment costs associated with preemption.…”

Fifer

“…•Public Cloud: There are several research works that optimize for the resource provisioning cost in the public cloud. These works broadly fall into two categories: (i) tuning the auto-scaling policy based on changing needs (e.g., Spot, On-Demand) [17,29,42,44,45,78,91], (ii) predicting peak loads and offering proactive provisioning based auto-scaling policy [42ś44, 63,79,97]. Fifer uses similar load prediction models and auto-scales containers but with respect to serverless function chains.…”

Minimizing Cost and Maximizing Performance for Cloud Platforms

“…There are several research works that optimize for the resource provisioning cost for various tenant applications in the public cloud. These works broadly fall into two categories: (i) tuning the auto-scaling policy based on changing needs [7,9,19,24,25,34,39], (ii) prediction-based proactive provisioning auto-scaling policy [19,23,24,28,35]. Complementary to these approaches, we propose to user serverless functions along with VMs for cost-effective autoscaling.…”