Spyre: A Resource Management Framework for Container-Based Clouds

Linux container technology is seeing rapid adoption in the last few years with its usage and enhancement as platform for building and deploying applications, for example, by Docker. Services are being stood up in public clouds that offer different frameworks for launching and managing user-built containers. At the same time cloud services are beginning to examine container-based deployment as possible alternative to or in conjunction with virtual machines for deploying those services. However, when multiple tenants’ containers get deployed on the same system there is little inherent isolation between tenants in existing containers-as-a-service platforms.

In the Spyre project we focus on developing a resource management framework that allows us to provide performance isolation between multiple tenants deploying containers in the cloud. We introduce the notion of a slice that provides the resource partition for one tenant within a system which com-prises of a defined set of resources—cores, memory, memory bandwidth, network bandwidth, storage, storage bandwidth etc. These are unique to that slice and disjoint from the resources commandeered for any other slice.

In this presentation, we socialize our concept of slices and discuss how they provide the performance isolation important to multi-tenant cloud services that care about tail latencies among other things. We discuss how we think they relate to concepts in existing container platforms and can be adopted for meeting performance-sensitive needs for cloud services. We then introduce our current implementation and invite audience to participate in a broader discussion on the challenges for performance isolation and management for container-based cloud frameworks.