On Thu, Nov 16, 2023 at 1:00 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > On Wed 15-11-23 17:33:51, Yafang Shao wrote: > > On Wed, Nov 15, 2023 at 4:45 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > > > On Wed 15-11-23 09:52:38, Yafang Shao wrote: > > > > On Wed, Nov 15, 2023 at 12:58 AM Casey Schaufler <casey@xxxxxxxxxxxxxxxx> wrote: > > > > > > > > > > On 11/14/2023 3:59 AM, Yafang Shao wrote: > > > > > > On Tue, Nov 14, 2023 at 6:15 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > > >> On Mon 13-11-23 11:15:06, Yafang Shao wrote: > > > > > >>> On Mon, Nov 13, 2023 at 12:45 AM Casey Schaufler <casey@xxxxxxxxxxxxxxxx> wrote: > > > > > >>>> On 11/11/2023 11:34 PM, Yafang Shao wrote: > > > > > >>>>> Background > > > > > >>>>> ========== > > > > > >>>>> > > > > > >>>>> In our containerized environment, we've identified unexpected OOM events > > > > > >>>>> where the OOM-killer terminates tasks despite having ample free memory. > > > > > >>>>> This anomaly is traced back to tasks within a container using mbind(2) to > > > > > >>>>> bind memory to a specific NUMA node. When the allocated memory on this node > > > > > >>>>> is exhausted, the OOM-killer, prioritizing tasks based on oom_score, > > > > > >>>>> indiscriminately kills tasks. This becomes more critical with guaranteed > > > > > >>>>> tasks (oom_score_adj: -998) aggravating the issue. > > > > > >>>> Is there some reason why you can't fix the callers of mbind(2)? > > > > > >>>> This looks like an user space configuration error rather than a > > > > > >>>> system security issue. > > > > > >>> It appears my initial description may have caused confusion. In this > > > > > >>> scenario, the caller is an unprivileged user lacking any capabilities. > > > > > >>> While a privileged user, such as root, experiencing this issue might > > > > > >>> indicate a user space configuration error, the concerning aspect is > > > > > >>> the potential for an unprivileged user to disrupt the system easily. > > > > > >>> If this is perceived as a misconfiguration, the question arises: What > > > > > >>> is the correct configuration to prevent an unprivileged user from > > > > > >>> utilizing mbind(2)?" > > > > > >> How is this any different than a non NUMA (mbind) situation? > > > > > > In a UMA system, each gigabyte of memory carries the same cost. > > > > > > Conversely, in a NUMA architecture, opting to confine processes within > > > > > > a specific NUMA node incurs additional costs. In the worst-case > > > > > > scenario, if all containers opt to bind their memory exclusively to > > > > > > specific nodes, it will result in significant memory wastage. > > > > > > > > > > That still sounds like you've misconfigured your containers such > > > > > that they expect to get more memory than is available, and that > > > > > they have more control over it than they really do. > > > > > > > > And again: What configuration method is suitable to limit user control > > > > over memory policy adjustments, besides the heavyweight seccomp > > > > approach? > > > > > > This really depends on the workloads. What is the reason mbind is used > > > in the first place? > > > > It can improve their performance. > > > > > Is it acceptable to partition the system so that > > > there is a numa node reserved for NUMA aware workloads? > > > > As highlighted in the commit log, our preference is to configure this > > memory policy through kubelet using cpuset.mems in the cpuset > > controller, rather than allowing individual users to set it > > independently. > > OK, I have missed that part. > > > > If not, have you > > > considered (already proposed numa=off)? > > > > The challenge at hand isn't solely about whether users should bind to > > a memory node or the deployment of workloads. What we're genuinely > > dealing with is the fact that users can bind to a specific node > > without our explicit agreement or authorization. > > mbind outside of the cpuset shouldn't be possible (policy_nodemask). So > if you are configuring cpusets already then mbind should add much to a > problem. I can see how you can have problems when you do not have any > NUMA partitioning in place because mixing NUMA aware and unaware > workloads doesn't really work out well when the memory is short on > supply. Right, we're trying to move NUMA aware workloads to dedicated servers. -- Regards Yafang
