A Performance Analysis on Auto Scaling Using Different Image Pulling Methods in Kubernetes
Korea Computer Congress (KCC 2023)
A scaling decision adds useful capacity only after the image is ready. We compared scale-up completion with and without pre-pulled images.
- Role
- First author · image readiness and autoscaling performance research
- Outcome
- For a 500MB image at HPA thresholds of 30%, 50%, and 70%, pre-pulling reduced scaling time by 38.21%, 36.46%, and 65.78%.
Background
When HPA scales pods up under load, if the target worker node lacks the container image it must pull it from a remote registry. When that pull coincides with a high-load scale-up, the overloaded state lasts longer. This paper quantifies how a node having the image pre-staged affects scale-up performance.
Experiment
A cluster of one master (Intel Core i9-10900K, DDR4 32 GB, Ubuntu 20.04) and three workers ran a php-apache service, driven by wget sending an HTTP request every 0.02 s while metrics were monitored for five minutes.
First, single-pod creation time was compared across 300 MB, 500 MB, and 1 GB images; then, with a 500 MB image and HPA CPU thresholds of 30 / 50 / 70, scale-up time and CPU usage were measured with the image present on only one of the three workers.
Results
Image pulling made up 72.8% of container start time on average when the image was local, and 98–99% when it was absent — growing with image size. When present, the pull took about 2.12 s regardless of size.
Without a local image, reaching the target replica count took over 60 seconds after scale-up was requested; with pre-pulling, it followed much sooner. For a 500MB image at CPU thresholds of 30%, 50%, and 70%, scaling time fell by 38.21%, 36.46%, and 65.78%, and CPU use by 39.53%, 31.11%, and 23.11%.
The important distinction is between requesting replicas and obtaining ready capacity. HPA determines replica count; node scheduling and image preparation are separate stages. A pod placed on a node without the image adds download time. The study proposed usage-aware image pre-pulling as follow-up work. I conducted this research as first author.