Hello, Waiman.
On Wed, Apr 12, 2023 at 11:37:53AM -0400, Waiman Long wrote:
> This patch series introduces a new "isolcpus" partition type to the
> existing list of {member, root, isolated} types. The primary reason
> of adding this new "isolcpus" partition is to facilitate the
> distribution of isolated CPUs down the cgroup v2 hierarchy.
>
> The other non-member partition types have the limitation that their
> parents have to be valid partitions too. It will be hard to create a
> partition a few layers down the hierarchy.
>
> It is relatively rare to have applications that require creation of
> a separate scheduling domain (root). However, it is more common to
> have applications that require the use of isolated CPUs (isolated),
> e.g. DPDK. One can use the "isolcpus" or "nohz_full" boot command options
> to get that statically. Of course, the "isolated" partition is another
> way to achieve that dynamically.
>
> Modern container orchestration tools like Kubernetes use the cgroup
> hierarchy to manage different containers. If a container needs to use
> isolated CPUs, it is hard to get those with existing set of cpuset
> partition types. With this patch series, a new "isolcpus" partition
> can be created to hold a set of isolated CPUs that can be pull into
> other "isolated" partitions.
>
> The "isolcpus" partition is special that there can have at most one
> instance of this in a system. It serves as a pool for isolated CPUs
> and cannot hold tasks or sub-cpusets underneath it. It is also not
> cpu-exclusive so that the isolated CPUs can be distributed down the
> sibling hierarchies, though those isolated CPUs will not be useable
> until the partition type becomes "isolated".
>
> Once isolated CPUs are needed in a cgroup, the administrator can write
> a list of isolated CPUs into its "cpuset.cpus" and change its partition
> type to "isolated" to pull in those isolated CPUs from the "isolcpus"
> partition and use them in that cgroup. That will make the distribution
> of isolated CPUs to cgroups that need them much easier.
I'm not sure about this. It feels really hacky in that it side-steps the
distribution hierarchy completely. I can imagine a non-isolated cpuset
wanting to allow isolated cpusets downstream but that should be done
hierarchically - e.g. by allowing a cgroup to express what isolated cpus are
allowed in the subtree. Also, can you give more details on the targeted use
cases?
Thanks.
--
tejun
