There are cases with workstealing where once a VT is stolen to make progress there is little point to keep it returning to its assigned scheduler (it can cause a cache miss!) and work stealing in the latest form could still redistribute it if the running one is overwhelmed, so it could be beneficial to keep it running where it is, relying on work stealing to do its job, and keeping the last CPI cache who handles it, hot.
The same won't apply to JDK pollers as they are really more tied to a specific scheduler, in their intention.
There are cases with workstealing where once a VT is stolen to make progress there is little point to keep it returning to its assigned scheduler (it can cause a cache miss!) and work stealing in the latest form could still redistribute it if the running one is overwhelmed, so it could be beneficial to keep it running where it is, relying on work stealing to do its job, and keeping the last CPI cache who handles it, hot.
The same won't apply to JDK pollers as they are really more tied to a specific scheduler, in their intention.