[ Picking this back up, I was out of the country last week. Note that we are also wrestling with some DMARC issues as it was just activated for Google.com so apologies if people do not receive this directly. ]
On Tue, Aug 25, 2015 at 2:02 PM, Tejun Heo <t...@kernel.org> wrote: > Hello, > > On Mon, Aug 24, 2015 at 04:06:39PM -0700, Paul Turner wrote: >> > This is an erratic behavior on cpuset's part tho. Nothing else >> > behaves this way and it's borderline buggy. >> >> It's actually the only sane possible interaction here. >> >> If you don't overwrite the masks you can no longer manage cpusets with >> a multi-threaded application. >> If you partially overwrite the masks you can create a host of >> inconsistent behaviors where an application suddenly loses >> parallelism. > > It's a layering problem. It'd be fine if cpuset either did "layer > per-thread affinities below w/ config change notification" or "ignore > and/or reject per-thread affinities". What we have now is two layers > manipulating the same field without any mechanism for coordination. > I think this is a mischaracterization. With respect to the two proposed solutions: a) Notifications do not solve this problem. b) Rejecting per-thread affinities is a non-starter. It's absolutely needed. (Aside: This would also wholly break the existing sched_setaffinity/getaffinity syscalls.) I do not think this is a layering problem. This is more like C++: there is no sane way to concurrently use all the features available, however, reasonably self-consistent subsets may be chosen. >> The *only* consistent way is to clobber all masks uniformly. Then >> either arrange for some notification to the application to re-sync, or >> use sub-sub-containers within the cpuset hierarchy to advertise >> finer-partitions. > > I don't get it. How is that the only consistent way? Why is making > irreversible changes even a good way? Just layer the masks and > trigger notification on changes. I'm not sure if you're agreeing or disagreeing here. Are you saying there's another consistent way from "clobbering then triggering notification changes" since it seems like that's what is rejected and then described. This certainly does not include any provisions for reversibility (which I think is a non-starter). With respect to layering: Are you proposing we maintain a separate mask for sched_setaffinity and cpusets, then do different things depending on their intersection, or lack thereof? I feel this would introduce more consistencies than it would solve as these masks would not be separately inspectable from user-space, leading to confusing interactions as they are changed. There are also very real challenges with how any notification is implemented, independent of delivery: The 'main' of an application often does not have good control or even understanding over its own threads since many may be library managed. Designation of responsibility for updating these masks is difficult. That said, I think a notification would still be a useful improvement here and that some applications would benefit. At the very least, I do not think that cpuset's behavior here can be dismissed as unreasonable. > >> I don't think the case of having a large compute farm with >> "unimportant" and "important" work is particularly fringe. Reducing >> the impact on the "important" work so that we can scavenge more cycles >> for the latency insensitive "unimportant" is very real. > > What if optimizing cache allocation across competing threads of a > process can yield, say, 3% gain across large compute farm? Is that > fringe? Frankly, yes. If you have a compute farm sufficiently dedicated to a single application I'd say that's a fairly specialized use. I see no reason why a more 'technical' API should be a challenge for such a user. The fundamental point here was that it's ok for the API of some controllers to be targeted at system rather than user control in terms of interface. This does not restrict their use by users where appropriate. > >> Right, but it's exactly because of _how bad_ those other mechanisms >> _are_ that cgroups was originally created. Its growth was not >> managed well from there, but let's not step away from the fact that >> this interface was created to solve this problem. > > Sure, at the same time, please don't forget that there are ample > reasons we can't replace more basic mechanisms with cgroups. I'm not > saying this can't be part of cgroup but rather that it's misguided to > do plunge into cgroups as the first and only step. > So there is definitely a proliferation of discussion regarding applying cgroups to other problems which I agree we need to take a step back and re-examine. However, here we're fundamentally talking about APIs designed to partition resources which is the problem that cgroups was introduced to address. If we want to introduce another API to do that below the process level we need to talk about why it's fundamentally different for processes versus threads, and whether we want two APIs that interface with the same underlying kernel mechanics. > More importantly, I am extremely doubtful that we understand the usage > scenarios and their benefits very well at this point and want to avoid > over-committing to something we'll look back and regret. As it > currently stands, this has a high likelihood of becoming a mismanaged > growth. I don't disagree with you with respect to new controllers, but I worry this is forking the discussion somewhat. There are two issues being conflated here: 1) The need for per-thread resource control and what such an API should look like. 2) The proliferation of new controllers, such as CAT. We should try to focus on (1) here as that is the most immediate for forward progress. We can certainly draw anecdotes from (2) but we do know (1) applies to existing controllers (e.g. cpu/cpuacct/cpuset). > > For the cache allocation thing, I'd strongly suggest something way > simpler and non-commmittal - e.g. create a char device node with > simple configuration and basic access control. If this *really* turns > out to be useful and its configuration complex enough to warrant > cgroup integration, let's do it then, and if we actually end up there, > I bet the interface that we'd come up with at that point would be > different from what people are proposing now. As above, I really want to focus on (1) so I will be brief here: Making it a char device requires yet-another adhoc method of describing process groupings that a configuration should apply to and yet-another set of rules for its inheritance. Once we merge it, we're committed to backwards support of the interface either way, I do not see what reimplementing things as a char device or sysfs or seqfile or other buys us over it being cgroupfs in this instance. I think that the real problem here is that stuff gets merged that does not follow the rules of how something implemented with cgroups must behave (typically respect with to a hierarchy); which is obviously increasingly incompatible with a model where we have a single hierarchy. But, provided that we can actually define those rules; I do not see the gain in denying the admission of new controller which is wholly consistent with them. It does not really measurably add to the complexity of the implementation (and it greatly reduces it where grouping semantics are desired). > >> > Yeah, I understand the similarity part but don't buy that the benefit >> > there is big enough to introduce a kernel API which is expected to be >> > used by individual programs which is radically different from how >> > processes / threads are organized and applications interact with the >> > kernel. >> >> Sorry, I don't quite follow, in what way is it radically different? >> What is magically different about a process versus a thread in this >> sub-division? > > I meant that cgroupfs as opposed to most other programming interfaces > that we publish to applications. We already have process / thread > hierarchy which is created through forking/cloning and conventions > built around them for interaction. I do not think the process/thread hierarchy is particularly comparable as it is both immutable and not a partition. It expresses resource parenting only. The only common operation performed on it is killing a sub-tree. > No sane application programming > interface requires individual applications to open a file somewhere, > echo some values to it and use directory operations to manage its > organization. Will get back to this later. > >> > All controllers only get what their ancestors can hand down to them. >> > That's basic hierarchical behavior. >> >> And many users want non work-conserving systems in which we can add >> and remove idle resources. This means that how much bandwidth an >> ancestor has is not fixed in stone. > > I'm having a hard time following you on this part of the discussion. > Can you give me an example? For example, when a system is otherwise idle we might choose to give an application additional memory or cpu resources. These may be reclaimed in the future, such an update requires updating children to be compatible with a parents' new limits. > >> > If that's the case and we fail miserably at creating a reasonable >> > programming interface for that, we can always revive thread >> > granularity. This is mostly a policy decision after all. >> >> These interfaces should be presented side-by-side. This is not a >> reasonable patch-later part of the interface as we depend on it today. > > Revival of thread affinity is trivial and will stay that way for a > long time and the transition is already gradual, so it'll be a lost > opportunity but there is quite a bit of maneuvering room. Anyways, on > with the sub-process interface. > > Skipping description of the problems with the current setup here as > I've repated it a couple times in this thread already. > > On the other sub-thread, I said that process/thread tree and cgroup > association are inherently tied. This is because a new child task is > always born into the parent's cgroup and the only reason cgroup works > on system management use cases is because system management often > controls enough of how processes are created. > > The flexible migration that cgroup supports may suggest that an > external agent with enough information can define and manage > sub-process hierarchy without involving the target application but > this doesn't necessarily work because such information is often only > available in the application itself and the internal thread hierarchy > should be agreeable to the hierarchy that's being imposed upon it - > when threads are dynamically created, different parts of the hierarchy > should be created by different parent thread. I think what's more important here is that you can define it to work. This does require cooperation between the external agent and the application in the layout of the application's hierarchy. But this is something we do use. A good example would be the surfacing of public and private cpus previously discussed to the application. > > Also, the problem with external and in-application manipulations > stepping on each other's toes is mostly not caused by individual > config settings but by the possibility that they may try to set up > different hierarchies or modify the existing one in a way which is not > expected by the other. How is this different from say signals or ptrace or any file-system modification? This does not seem a problem inherent to cgroups. > > Given that thread hierarchy already needs to be compatible with > resource hierarchy, is something unix programs already understands and > thus can render itself to an a lot more conventional interface which > doesn't cause organizational conflicts, I think it's logical to use > that for sub-process resource distribution. > > So, it comes down to sth like the following > > set_resource($TID, $FLAGS, $KEY, $VAL) > > - If $TID isn't already a resource group leader, it creates a > sub-cgroup, sets $KEY to $VAL and moves $PID and all its descendants > to it. > > - If $TID is already a resource group leader, set $KEY to $VAL. > > - If the process is moved to another cgroup, the sub-hierarchy is > preserved. > Honestly, I find this API awkward: 1) It depends on "anchor" threads to define groupings. 2) It does not allow thread-level hierarchies to be created 3) When coordination with an external agent is desired this defines no common interface that can be shared. Directories are an extremely useful container. Are you proposing applications would need to somehow publish the list of anchor-threads from (1)? What if I want to set up state that an application will attaches threads to [consider cpuset example above]? 4) How is the cgroup property to $KEY translation defined? This feels like an ioctl and no more natural than the file-system. It also does not seem to resolve your concerns regarding races; the application must still coordinate internally when concurrently calling set_resource(). 5) How does an external agent coordinate when a resource must be removed from a sub-hierarchy? On a larger scale, what properties do you feel this separate API provides that would not be also supported by instead exposing sub-process hierarchies via /proc/self or similar. Perhaps it would help to enumerate the the key problems we're trying to solve with the choice of this interface? 1) Thread spanning trees within the cgroup hierarchy. (Immediately fixed, only processes are present on the cgroup-mount) 2) Interactions with the parent process moving within the hierarchy 3) Potentially supporting move operations within a hierarchy Are there other cruxes? > The reality is a bit more complex and cgroup core would need to handle > implicit leaf cgroups and duplicating sub-hierarchy. The biggest > complexity would be extending atomic multi-thread migrations to > accomodate multiple targets but it already does atomic multi-task > migrations and performing the migrations back-to-back should work. > Controller side changes wouldn't be much. Copying configs to clone > sub-hierarchy and specifying which are availble should be about it. > > This should give applications a simple and straight-forward interface > to program against while avoiding all the issues with exposing > cgroupfs directly to individual applications. Is your primary concern here (2) above? E.g. that moving the parent process means that the location we write to for sub-process updates is not consistent? Or other? For issues involving synchronization, what's proposed at least feels no different to what we face today. > >> > So, the proposed patches already merge cpu and cpuacct, at least in >> > appearance. Given the kitchen-sink nature of cpuset, I don't think it >> > makes sense to fuse it with cpu. >> >> Arguments in favor of this: >> a) Today the load-balancer has _no_ understanding of group level >> cpu-affinity masks. >> b) With SCHED_NUMA, we can benefit from also being able to apply (b) >> to understand which nodes are usable. > > Controllers can cooperate with each other on the unified hierarchy - > cpu can just query the matching cpuset css about the relevant > attributes and the results will always be properly hierarchical for > cpu too. There's no reason to merge the two controllers for that Let's shelve this for now. > > Thanks. > > -- > tejun -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/