Anyone have ideas about what to do when using cgroup weights rather than max?
I'm currently in the process of removing cpu.max from our clusters, to allow applications to better utilize the available cpu time which currently is just being wasted. We will use cpu weights to make sure that cpu time is fairly allocated during contention, and to not oversubscribe the hosts, but I'm sure that among the thousands of applications that are running on those clusters today, many will be relying on cpu.max to size threadpools etc.
On the one hand, we do want applications to use the available cpu time, but at the same time they need to not kill themselves by running out of memory.
24 comments
I'm currently in the process of removing cpu.max from our clusters, to allow applications to better utilize the available cpu time which currently is just being wasted. We will use cpu weights to make sure that cpu time is fairly allocated during contention, and to not oversubscribe the hosts, but I'm sure that among the thousands of applications that are running on those clusters today, many will be relying on cpu.max to size threadpools etc.
On the one hand, we do want applications to use the available cpu time, but at the same time they need to not kill themselves by running out of memory.