Melding the Serverless Control Plane with the Conventional Cluster Manager for Speed and Resource Efficiency

TL;DR

PulseNet tackles the core serverless challenge of achieving low-latency scaling without sacrificing compatibility with conventional cluster managers. It introduces a dual-track control plane: a standard track using a mature cluster manager for sustainable traffic, and an expedited track with node-local agents that rapidly instantiate disposable Emergency Instances for bursts, bypassing the main cluster manager. Evaluation on production-like traces shows PulseNet delivers 1.5–3.5x performance improvements over Kubernetes-compatible systems and up to 70% cost reductions, while maintaining compatibility and reducing memory waste compared to FaaS-specialized designs. The work demonstrates that by exploiting bimodal traffic characteristics and separating fast burst handling from steady-state management, serverless platforms can achieve high performance and efficiency at scale, enabling tighter co-location of FaaS and BaaS components within a single cluster.

Abstract

Serverless platforms face a trade-off: conventional cluster managers like Kubernetes offer compatibility for co-locating Function-as-a-Service (FaaS) and Backend-as-a-Service (BaaS) components of serverless applications, at the cost of high cold-start latency, whereas specialized FaaS-only systems like Dirigent achieve low latency by sacrificing compatibility, preventing integrated management and optimization. Our analysis reveals that FaaS traffic is bimodal: predictable, sustainable traffic consumes >98% of cluster resources, whereas sporadic, excessive bursts stress the control plane's scaling latency, not its throughput. With these insights, we design PulseNet, a serverless architecture that uses a dual-track control plane tailored to both traffic types. PulseNet's standard track manages sustainable traffic with long-lived, full-featured Regular Instances under a conventional cluster manager, preserving compatibility for the majority of the workload. To handle excessive traffic, an expedited track bypasses the slow manager to rapidly create short-lived, disposable Emergency Instances, minimizing cold-start latency and resource waste from idle instances. This hybrid approach achieves 35% better performance than Dirigent, a FaaS-only system, on a production workload at the same cost and outperforms other Kubernetes-compatible systems by 1.5-3.5x, reducing the cost by up to 70%.

Figures (13)

• Figure 1: Instance number scaling over time in response to the changes in the in-flight request concurrency in the state-of-the-art systems and PulseNet. Kn scales too slowly, whereas Kn-Sync incurs high costs by keeping instances idle for prolonged periods. Dirigent's behavior is similar to Kn (not shown).
• Figure 2: High-level overview of serverless architecture. The cluster manager is on the path only for cold invocations.
• Figure 3: Cumulative distribution functions (CDFs) for the components of delays occurring in the systems with synchronous (Kn-Sync) and asynchronous (Kn) control planes.
• Figure 4: Instance creation time breakdown in Knative.
• Figure 5: The delays occurring in the Knative control plane under various instance-creation rates, measured with a microbenchmark. The red and black lines show the required instance creation rates at the 50-th and 99-th percentiles, respectively, when replaying invocations from a sampled production trace in simulated synchronous (Kn-Sync) and asynchronous (Kn) control planes.