The article explores the issue of CPU contention in virtualized environments, particularly focusing on how the hypervisor's scheduling of virtual CPUs can lead to high wait times for processes. It cites a practical experience with a RabbitMQ cluster encountering high steal time percentages, indicating a contention issue. The author details a simulation setup on AWS to demonstrate CPU contention with burstable instances and emphasizes the importance of monitoring thread wait times to detect and resolve contention effectively.
Ruin is the destination toward which all men rush, each pursuing his own best interest in a society that believes in the freedom of the commons.
The CFS scheduler sees only virtual CPUs. The hypervisor is juggling those vCPUs onto a finite set of cores. When everyone wants to run at once, somebody waits.
Burstable instances have a baseline CPU but can burst higher using CPU credits. It's important to run stress-ng long enough before continuing to expend CPU credits.
Detecting contention requires monitoring how long a runnable thread waits.
Collection
[
|
...
]