Hi, David, Thank you very much for your attention. >The problem you're trying to solve only exists in complex graphs with >different per-vertex parallelism. If the parallelism is set globally >(assuming the pipeline has roughly even data skew), the algorithm could >make things slightly worse by eliminating some local exchanges. Is that >correct?
Your understanding is accurate, and it's undeniable that such use case scenarios exist. >Where I'm headed with this is that there could be a hybrid strategy that >provides a reasonable default when the pipeline uses slot-sharing (for >per-vertex parallelism, use the new strategy; for global parallelism use >the old one). It's always a shame if improvements like this end up being a >power-user feature and very few workloads benefit from it. Any thoughts? The concept of letting the engine determine the scheduling strategy based on a predefined rule is excellent. This approach aims to maximize job performance while minimizing user intervention. It might not need to rush into implementing this rule at this moment. What I mean is, we can evaluate and develop a well-founded rule in future work. Nonetheless, we can still consider this rule in advance so that it can be validated after the feature's release. Additionally, if we decide to implement this rule in the future, it should be introduced as a switch. As you pointed out, we currently don't take data characteristics' impact on task resource allocation in the actual environment into account. Therefore, implementing it as a switch will offer users greater flexibility. Of course, it will add a little complexity to users' understanding of this parameter. I'm also eager to hear from other contributors regarding it and looking forward to your reply. Best, Yuepeng. On 2023/10/02 20:37:12 David Morávek wrote: > Hello Yuepeng, > > The FLIP reads sane; nice work! To re-phrase my understanding: > > The problem you're trying to solve only exists in complex graphs with > different per-vertex parallelism. If the parallelism is set globally > (assuming the pipeline has roughly even data skew), the algorithm could > make things slightly worse by eliminating some local exchanges. Is that > correct? > > Where I'm headed with this is that there could be a hybrid strategy that > provides a reasonable default when the pipeline uses slot-sharing (for > per-vertex parallelism, use the new strategy; for global parallelism use > the old one). It's always a shame if improvements like this end up being a > power-user feature and very few workloads benefit from it. Any thoughts? > > Best, > D. > > On Sun, Oct 1, 2023 at 1:38 PM Yangze Guo <karma...@gmail.com> wrote: > > > Hi, Rui, > > > > 1. With the current mechanism, when physical slots are offered from > > TM, the JobMaster will start deploying tasks and synchronizing their > > states. With the addition of the waiting mechanism, IIUC, the > > JobMaster will deploy and synchronize the states of all tasks only > > after all resources are available. The task deployment and state > > synchronization both occupy the JobMaster's RPC main thread. In > > complex jobs with a lot of tasks, this waiting mechanism may increase > > the pressure on the JobMaster and increase the end-to-end job > > deployment time. > > > > 2. From my understanding, if user enable the > > cluster.evenly-spread-out-slots, > > LeastUtilizationResourceMatchingStrategy will be used to determine the > > slot distribution and the slot allocation in the three TM will be > > (taskmanager.numberOfTaskSlots=3): > > TM1: 3 slot > > TM2: 2 slot > > TM3: 2 slot > > > > Best, > > Yangze Guo > > > > On Sun, Oct 1, 2023 at 6:14 PM Rui Fan <1996fan...@gmail.com> wrote: > > > > > > Hi Shammon, > > > > > > Thanks for your feedback as well! > > > > > > > IIUC, the overall balance is divided into two parts: slot to TM and > > task > > > to slot. > > > > 1. Slot to TM is guaranteed by SlotManager in ResourceManager > > > > 2. Task to slot is guaranteed by the slot pool in JM > > > > > > > > These two are completely independent, what are the benefits of unifying > > > > these two into one option? Also, do we want to share the same > > > > option between SlotPool in JM and SlotManager in RM? This sounds a bit > > > > strange. > > > > > > Your understanding is totally right, the balance needs 2 parts: slot to > > TM > > > and task to slot. > > > > > > As I understand, the following are benefits of unifying them into one > > > option: > > > > > > - Flink users don't care about these principles inside of flink, they > > don't > > > know these 2 parts. > > > - If flink provides 2 options, flink users need to set 2 options for > > their > > > job. > > > - If one option is missed, the final result may not be good. (Users may > > > have questions when using) > > > - If flink just provides 1 option, enabling one option is enough. (Reduce > > > the probability of misconfiguration) > > > > > > Also, Flink’s options are user-oriented. Each option represents a switch > > or > > > parameter of a feature. > > > A feature may be composed of multiple components inside Flink. > > > It might be better to keep only one switch per feature. > > > > > > Actually, the cluster.evenly-spread-out-slots option is used between > > > SlotPool in JM and SlotManager in RM. 2 components to ensure > > > this feature works well. > > > > > > Please correct me if my understanding is wrong, > > > and looking forward to your feedback, thanks! > > > > > > Best, > > > Rui > > > > > > On Sun, Oct 1, 2023 at 5:52 PM Rui Fan <1996fan...@gmail.com> wrote: > > > > > > > Hi Yangze, > > > > > > > > Thanks for your feedback! > > > > > > > > > 1. Is it possible for the SlotPool to get the slot allocation results > > > > > from the SlotManager in advance instead of waiting for the actual > > > > > physical slots to be registered, and perform pre-allocation? The > > > > > benefit of doing this is to make the task deployment process > > smoother, > > > > > especially when there are a large number of tasks in the job. > > > > > > > > Could you elaborate on that? I didn't understand what's the benefit and > > > > smoother. > > > > > > > > > 2. If user enable the cluster.evenly-spread-out-slots, the issue in > > > > > example 2 of section 2.2.3 can be resolved. Do I understand it > > > > > correctly? > > > > > > > > The example assigned result is the final allocation result when flink > > > > user enables the cluster.evenly-spread-out-slots. We think the > > > > assigned result is expected, so I think your understanding is right. > > > > > > > > Best, > > > > Rui > > > > > > > > On Thu, Sep 28, 2023 at 1:10 PM Shammon FY <zjur...@gmail.com> wrote: > > > > > > > >> Thanks Yuepeng for initiating this discussion. > > > >> > > > >> +1 in general too, in fact we have implemented a similar mechanism > > > >> internally to ensure a balanced allocation of tasks to slots, it works > > > >> well. > > > >> > > > >> Some comments about the mechanism > > > >> > > > >> 1. This mechanism will be only supported in `SlotPool` or both > > `SlotPool` > > > >> and `DeclarativeSlotPool`? Currently the two slot pools are used in > > > >> different schedulers. I think this will also bring value to > > > >> `DeclarativeSlotPool`, but currently FLIP content seems to be based on > > > >> `SlotPool`, right? > > > >> > > > >> 2. In fine-grained resource management, we can set different resource > > > >> requirements for different nodes, which means that the resources of > > each > > > >> slot are different. What should be done when the slot selected by the > > > >> round-robin strategy cannot meet the resource requirements? Will this > > lead > > > >> to the failure of the balance strategy? > > > >> > > > >> 3. Is the assignment of tasks to slots balanced based on region or job > > > >> level? When multiple TMs fail over, will it cause the balancing > > strategy > > > >> to > > > >> fail or even worse? What is the current processing strategy? > > > >> > > > >> For Zhuzhu and Rui: > > > >> > > > >> IIUC, the overall balance is divided into two parts: slot to TM and > > task > > > >> to > > > >> slot. > > > >> 1. Slot to TM is guaranteed by SlotManager in ResourceManager > > > >> 2. Task to slot is guaranteed by the slot pool in JM > > > >> > > > >> These two are completely independent, what are the benefits of > > unifying > > > >> these two into one option? Also, do we want to share the same > > > >> option between SlotPool in JM and SlotManager in RM? This sounds a bit > > > >> strange. > > > >> > > > >> Best, > > > >> Shammon FY > > > >> > > > >> > > > >> > > > >> On Thu, Sep 28, 2023 at 12:08 PM Rui Fan <1996fan...@gmail.com> > > wrote: > > > >> > > > >> > Hi Zhu Zhu, > > > >> > > > > >> > Thanks for your feedback here! > > > >> > > > > >> > You are right, user needs to set 2 options: > > > >> > - cluster.evenly-spread-out-slots=true > > > >> > - slot.sharing-strategy=TASK_BALANCED_PREFERRED > > > >> > > > > >> > Update it to one option is useful at user side, so > > > >> > `taskmanager.load-balance.mode` sounds good to me. > > > >> > I want to check some points and behaviors about this option: > > > >> > > > > >> > 1. The default value is None, right? > > > >> > 2. When it's set to Tasks, how to assign slots to TM? > > > >> > - Option1: It's just check task number > > > >> > - Option2: It''s check the slot number first, then check the > > > >> > task number when the slot number is the same. > > > >> > > > > >> > Giving an example to explain what's the difference between them: > > > >> > > > > >> > - A session cluster has 2 flink jobs, they are jobA and jobB > > > >> > - Each TM has 4 slots. > > > >> > - The task number of one slot of jobA is 3 > > > >> > - The task number of one slot of jobB is 1 > > > >> > - We have 2 TaskManagers: > > > >> > - tm1 runs 3 slots of jobB, so tm1 runs 3 tasks > > > >> > - tm2 runs 1 slot of jobA, and 1 slot of jobB, so tm2 runs 4 > > tasks. > > > >> > > > > >> > Now, we need to run a new slot, which tm should offer it? > > > >> > - Option1: If we just check the task number, the tm1 is better. > > > >> > - Option2: If we check the slot number first, and then check task, > > the > > > >> tm2 > > > >> > is better > > > >> > > > > >> > The original FLIP selected option2, that's why we didn't add the > > > >> > third option. The option2 didn't break the semantics when > > > >> > `cluster.evenly-spread-out-slots` is true, and it just improve the > > > >> > behavior without the semantics is changed. > > > >> > > > > >> > In the other hands, if we choose option2, when user set > > > >> > `taskmanager.load-balance.mode` is Tasks. It also can achieve > > > >> > the goal when it's Slots. > > > >> > > > > >> > So I think the `Slots` enum isn't needed if we choose option2. > > > >> > Of course, If we choose the option1, the enum is needed. > > > >> > > > > >> > Looking forward to your feedback, thanks~ > > > >> > > > > >> > Best, > > > >> > Rui > > > >> > > > > >> > On Wed, Sep 27, 2023 at 9:11 PM Zhu Zhu <reed...@gmail.com> wrote: > > > >> > > > > >> > > Thanks Yuepeng and Rui for creating this FLIP. > > > >> > > > > > >> > > +1 in general > > > >> > > The idea is straight forward: best-effort gather all the slot > > requests > > > >> > > and offered slots to form an overview before assigning slots, > > trying > > > >> to > > > >> > > balance the loads of task managers when assigning slots. > > > >> > > > > > >> > > I have one comment regarding the configuration for ease of use: > > > >> > > > > > >> > > IIUC, this FLIP uses an existing config > > > >> 'cluster.evenly-spread-out-slots' > > > >> > > as the main switch of the new feature. That is, from user > > perspective, > > > >> > > with this improvement, the 'cluster.evenly-spread-out-slots' > > feature > > > >> not > > > >> > > only balances the number of slots on task managers, but also > > balances > > > >> the > > > >> > > number of tasks. This is a behavior change anyway. Besides that, > > it > > > >> also > > > >> > > requires users to set 'slot.sharing-strategy' to > > > >> > 'TASK_BALANCED_PREFERRED' > > > >> > > to balance the tasks in each slot. > > > >> > > > > > >> > > I think we can introduce a new config option > > > >> > > `taskmanager.load-balance.mode`, > > > >> > > which accepts "None"/"Slots"/"Tasks". > > > >> `cluster.evenly-spread-out-slots` > > > >> > > can be superseded by the "Slots" mode and get deprecated. In the > > > >> future > > > >> > > it can support more mode, e.g. "CpuCores", to work better for jobs > > > >> with > > > >> > > fine-grained resources. The proposed config option > > > >> > > `slot.request.max-interval` > > > >> > > then can be renamed to > > > >> > > `taskmanager.load-balance.request-stablizing-timeout` > > > >> > > to show its relation with the feature. The proposed > > > >> > `slot.sharing-strategy` > > > >> > > is not needed, because the configured "Tasks" mode will do the > > work. > > > >> > > > > > >> > > WDYT? > > > >> > > > > > >> > > Thanks, > > > >> > > Zhu Zhu > > > >> > > > > > >> > > Yuepeng Pan <panyuep...@apache.org> 于2023年9月25日周一 16:26写道： > > > >> > > > > > >> > >> Hi all, > > > >> > >> > > > >> > >> > > > >> > >> I and Fan Rui(CC’ed) created the FLIP-370[1] to support balanced > > > >> tasks > > > >> > >> scheduling. > > > >> > >> > > > >> > >> > > > >> > >> The current strategy of Flink to deploy tasks sometimes leads > > some > > > >> > >> TMs(TaskManagers) to have more tasks while others have fewer > > tasks, > > > >> > >> resulting in excessive resource utilization at some TMs that > > contain > > > >> > more > > > >> > >> tasks and becoming a bottleneck for the entire job processing. > > > >> > Developing > > > >> > >> strategies to achieve task load balancing for TMs and reducing > > job > > > >> > >> bottlenecks becomes very meaningful. > > > >> > >> > > > >> > >> > > > >> > >> The raw design and discussions could be found in the Flink > > JIRA[2] > > > >> and > > > >> > >> Google doc[3]. We really appreciate Zhu Zhu(CC’ed) for providing > > some > > > >> > >> valuable help and suggestions in advance. > > > >> > >> > > > >> > >> > > > >> > >> Please refer to the FLIP[1] document for more details about the > > > >> proposed > > > >> > >> design and implementation. We welcome any feedback and opinions > > on > > > >> this > > > >> > >> proposal. > > > >> > >> > > > >> > >> > > > >> > >> [1] > > > >> > >> > > > >> > > > > >> > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-370%3A+Support+Balanced+Tasks+Scheduling > > > >> > >> > > > >> > >> [2] https://issues.apache.org/jira/browse/FLINK-31757 > > > >> > >> > > > >> > >> [3] > > > >> > >> > > > >> > > > > >> > > https://docs.google.com/document/d/14WhrSNGBdcsRl3IK7CZO-RaZ5KXU2X1dWqxPEFr3iS8 > > > >> > >> > > > >> > >> > > > >> > >> Best, > > > >> > >> > > > >> > >> Yuepeng Pan > > > >> > >> > > > >> > > > > > >> > > > > >> > > > > > > >
