The Power of Clouds: The Power Metric for Cloud Auto-Scaling Under Attack

Abstract

Modern cloud applications are composed of networks of distributed microservices that process user traffic. Unlike static queueing networks, cloud applications are inherently dynamic: service capacity can be adjusted through auto-scaling in response to traffic load surges and attacks, often within a short time. Prior work on worst-case attacks in queueing networks has focused mainly on static settings where the defender may react to attacks, at most, through traffic re-routing over fixed resources. We propose to study dynamic auto-scaled services through Kleinrock's Power metric [3], which captures the tradeoff between delay and throughput. The metric is closely aligned with the goal of auto-scaling: maintaining sufficient capacity to sustain performance while avoiding unnecessary over-provisioning of resources. Using Power-optimality analysis, we study the economic impact of overload attacks. While auto-scaling improves the system's ability to preserve latency under attack, it comes at the cost of provisioning additional resources. Using two simple network structures, we show how dynamic capacity control changes the damage-maximizing strategy of the attacker, compared to un-scaled static systems.