Using Burstable Instances in the Public Cloud

Abstract

Amazon EC2 and Google Compute Engine (GCE) have recently introduced a new class of virtual machines called "burstable" instances that are cheaper than even the smallest traditional/regular instances. These lower prices come with reduced average capacity and increased variance. Using measurements from both EC2 and GCE, we identify key idiosyncrasies of resource capacity dynamism for burstable instances that set them apart from other instance types. Most importantly, certain resources for these instances appear to be regulated by deterministic token bucket like mechanisms. We find widely different types of disclosures by providers of the parameters governing these regulation mechanisms: full disclosure (e.g., CPU capacity for EC2 t2 instances), partial disclosure (e.g., CPU capacity and remote disk IO bandwidth for GCE shared-core instances), or no disclosure (network bandwidth for EC2 t2 instances). A tenant modeling these variations as random phenomena (as some recent work suggests) might make sub-optimal procurement and operation decisions. We present modeling techniques for a tenant to infer the properties of these regulation mechanisms via simple offline measurements. We also present two case studies of how certain memcached workloads might benefit from our modeling when operating on EC2 by: (i) augmenting cheap but low availability in-memory storage offered by spot instances with backup of popular content on burstable instances, and (ii) temporal multiplexing of multiple burstable instances to achieve the CPU or network bandwidth (and thereby throughput) equivalent of a more expensive regular EC2 instance.