To explain what
@emoga suggested:
If a project is configured to 0 % resource share, the work request policy of the client changes qualitatively at that project. It will emit work requests
only when the corresponding resource (CPU or GPU) completely runs out of work. (All 'runnable' work across all projects which are active on the client combined.) And it will request very little work, such that the server typically¹ sends only one new task per CPU or GPU, regardless of the client's work buffer size setting.
________
¹) I have seen more than one task per GPU received occasionally when Amicable Numbers was at 0 % resource share. I have never seen it receive more than 1 task per idle logical CPU on any CPU-only project with 0 % resource share.