[lkml]   [2020]   [Feb]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH v2 3/3] mm: memcontrol: recursive memory.low protection
Hi Michal,

did you have any more thoughts on this?

On Mon, Feb 03, 2020 at 04:52:02PM -0500, Johannes Weiner wrote:
> On Thu, Jan 30, 2020 at 06:00:20PM +0100, Michal Hocko wrote:
> > On Thu 19-12-19 15:07:18, Johannes Weiner wrote:
> > > Right now, the effective protection of any given cgroup is capped by
> > > its own explicit memory.low setting, regardless of what the parent
> > > says. The reasons for this are mostly historical and ease of
> > > implementation: to make delegation of memory.low safe, effective
> > > protection is the min() of all memory.low up the tree.
> > >
> > > Unfortunately, this limitation makes it impossible to protect an
> > > entire subtree from another without forcing the user to make explicit
> > > protection allocations all the way to the leaf cgroups - something
> > > that is highly undesirable in real life scenarios.
> > >
> > > Consider memory in a data center host. At the cgroup top level, we
> > > have a distinction between system management software and the actual
> > > workload the system is executing. Both branches are further subdivided
> > > into individual services, job components etc.
> > >
> > > We want to protect the workload as a whole from the system management
> > > software, but that doesn't mean we want to protect and prioritize
> > > individual workload wrt each other. Their memory demand can vary over
> > > time, and we'd want the VM to simply cache the hottest data within the
> > > workload subtree. Yet, the current memory.low limitations force us to
> > > allocate a fixed amount of protection to each workload component in
> > > order to get protection from system management software in
> > > general. This results in very inefficient resource distribution.
> >
> > I do agree that configuring the reclaim protection is not an easy task.
> > Especially in a deeper reclaim hierarchy. systemd tends to create a deep
> > and commonly shared subtrees. So having a protected workload really
> > requires to be put directly into a new first level cgroup in practice
> > AFAICT. That is a simpler example though. Just imagine you want to
> > protect a certain user slice.
> Can you elaborate a bit on this? I don't quite understand the two
> usecases you are contrasting here.
> > You seem to be facing a different problem though IIUC. You know how much
> > memory you want to protect and you do not have to care about the cgroup
> > hierarchy up but you do not know/care how to distribute that protection
> > among workloads running under that protection. I agree that this is a
> > reasonable usecase.
> I'm not sure I'm parsing this right, but the use case is this:
> When I'm running a multi-component workload on a host without any
> cgrouping, the individual components compete over the host's memory
> based on rate of allocation, how often they reference their memory and
> so forth. It's a need-based distribution of pages, and the weight can
> change as demand changes throughout the life of the workload.
> If I now stick several of such workloads into a containerized
> environment, I want to use memory.low to assign each workload as a
> whole a chunk of memory it can use - I don't want to assign fixed-size
> subchunks to each individual component of each workload! I want the
> same free competition *within* the workload while setting clear rules
> for competition *between* the different workloads.
> [ What I can do today to achieve this is disable the memory controller
> for the subgroups. When I do this, all pages of the workload are on
> one single LRU that is protected by one single memory.low.
> But obviously I lose any detailed accounting as well.
> This patch allows me to have the same recursive protection semantics
> while retaining accounting. ]
> > Those both problems however show that we have a more general
> > configurability problem for both leaf and intermediate nodes. They are
> > both a result of strong requirements imposed by delegation as you have
> > noted above. I am thinking didn't we just go too rigid here?
> The requirement for delegation is that child groups cannot claim more
> than the parent affords. Is that the restriction you are referring to?
> > Delegation points are certainly a security boundary and they should
> > be treated like that but do we really need a strong containment when
> > the reclaim protection is under admin full control? Does the admin
> > really have to reconfigure a large part of the hierarchy to protect a
> > particular subtree?
> >
> > I do not have a great answer on how to implement this unfortunately. The
> > best I could come up with was to add a "$inherited_protection" magic
> > value to distinguish from an explicit >=0 protection. What's the
> > difference? $inherited_protection would be a default and it would always
> > refer to the closest explicit protection up the hierarchy (with 0 as a
> > default if there is none defined).
> > A
> > / \
> > B C (low=10G)
> > / \
> > D E (low = 5G)
> >
> > A, B don't get any protection (low=0). C gets protection (10G) and
> > distributes the pressure to D, E when in excess. D inherits (low=10G)
> > and E overrides the protection to 5G.
> >
> > That would help both usecases AFAICS while the delegation should be
> > still possible (configure the delegation point with an explicit
> > value). I have very likely not thought that through completely. Does
> > that sound like a completely insane idea?
> >
> > Or do you think that the two usecases are simply impossible to handle
> > at the same time?
> Doesn't my patch accomplish this?
> Any cgroup or group of cgroups still cannot claim more than the
> ancestral protection for the subtree. If a cgroup says 10G, the sum of
> all children's protection will never exceed that. This ensures
> delegation is safe.
> But once an ancestor does declare protection, that value automatically
> applies recursively to the entire subtree. The subtree can *chose* to
> assign fixed shares of that to its cgroups. But it doesn't have to.
> While designing this, I also thought about a new magic token for the
> memory.low to distinguish "inherit" from 0 (or other values smaller
> than what the parent affords). But the only thing it buys you is that
> you allow children to *opt out* of protection the parent prescribes.
> And I'm not sure that's a valid usecase: If the parent specifies that
> a subtree shall receive a protection of 10G compared to other subtrees
> at that level, does it really make sense for children to opt out and
> reduce the subtree's protection?
> IMO, no. A cgroup should be able to influence competition among its
> children, it should not be able to change the rules of competition
> among its ancestors.

 \ /
  Last update: 2020-02-10 16:23    [W:0.143 / U:0.396 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site