Kubernetes node swap tests find gains in workload density
Tests using fast solid-state storage found density gains of up to three times, often with little or no latency cost.
AI illustrationTests of Kubernetes nodes using swap backed by fast NVMe solid state drives found workload density gains of up to three times, often with little or no latency cost. Support for running nodes with swap enabled reached General Availability in v1.34.
The benchmark work examined three types of workload: CI/CD kernel builds, sandboxed headless browsers and isolated Python runtimes. Its authors used swap to move dormant memory out of RAM, allowing a node to accommodate more pods. The reported gains describe those tests and do not establish the same improvement for every workload.
Memory can become a limiting resource before processing capacity is exhausted. Agentic AI workloads add to that pressure because starting them and running untrusted code can require a large memory footprint. They can then remain resident while waiting for another prompt, consuming memory even during idle periods.
The approach described in the tests addresses that idle memory by placing it on fast storage. The authors found that this could increase the number of pods supported by a node, with the size of the benefit depending on the tested workload.