THIS IS A TEST INSTANCE. ALL YOUR CHANGES WILL BE LOST!!!!
Status
Current state: Under Discussion
...
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
Currently (Flink 1.9), Flink adopts a coarse grained resource management approach, where tasks are deployed into as many as the job’s max parallelism of predefined slots, regardless of how much resource each task / operator can use.
...
- Tasks may have different parallelisms, thus not all of the slots contains an entire pipeline of tasks. For slots with less tasks, the slot resource predefined for an entire pipeline might be wasteful.
- It could be hard to align slot resource with tasks requirements in all resource aspects (heap, network, managed, etc.).
In this FLIP, we We propose fine grained resource management, which optimizes resource utility in conditions where resource requirements of individual tasks are known or can be tuned.
...
- It works well for both Streaming and Batch jobs.
- It works well whether tasks’ resource requirements are specified or unknown.
This FLIP focuses on the operator resource management aspect of fine grained resource management.
Public Interfaces
- ResourceSpec (See Resource Requirements)
- Introduce new configuration "taskmanager.defaultSlotResourceFraction", while deprecate but stay compatible with the configuration "taskmanager.numberOfTaskSlots". (See Unknown Resource Requirements)RestAPI / WebUI (Need to adapt the RestAPI and WebUI to the dynamic slot model.)
Proposed Changes
Operator Resource
...
Requirements
Tasks can consume the following resources of task executors (based on FLIP-49 [1]):
...
For the first version, we do not support mixing operators with specified / unknown resource requirements in the same job. Either all or none of the operators of the same job should specify their resource requirements. StreamGraphGenerator should check this and throw an error when mixing of specified / unknown resource requirements is detected, during the compilation stage.
Managed Memory Allocation
Fraction Based Quota
Operators should not assume any absolute quota on managed memory. Instead, a relative quota should be applied, to unify memory management for both operators with specified and unknown resource requirements.
...
When the task is deployed to the task executor, operators should register their fractions to the memory manager before consuming any managed memory. The registration should return the absolute quota given the relative fraction. In this way, an operator can either consume managed memory respecting to its quota and assume the memory can be guaranteed, or leave it to the memory manager to limit its memory consumption and live with the possibility that allocating new memory may not always succeed.
Over-allocation and Revocation
For the first version, we do not allow operators to consume managed memory more than their quota.
In the future, we want to allow the operators to leverage the opportunistic available managed memory in the task executor. Operators may allocate managed memory more than their quota, as long as there are enough available managed memory in the task executor and the over-allocated memory can be revoked when needed by another task whose quota is not exceeded.
Slot Sharing
During the compiling stage, the StreamGraphGenerator first identifies pipelined regions in the job graph. A pipelined region is defined as the subset of vertices connected by pipelined edges in the job graph, which should always be scheduled together. Otherwise there might be a deadlock when downstream tasks cannot be scheduled due to lack of resources, while the upstream tasks cannot finish releasing the resources because no downstream tasks read the outputs.
...
- For tasks with specified resource requirements, we add up resource requirements of all the tasks in the slot sharing group, and request a slot with the sum resources.
- For tasks with unknown resource requirements, we request a slot with default resources.
Dynamic Slot Allocation
Dynamic Slot Model
Currently (Flink 1.9), a task executor contains a fixed number of slots, whose resource are predefined with total task executor resource and number of slots per task executor. These slots share the same life span as the task executor does. Slots are initially free, and are assigned to and freed by job masters.
...
- resources
...
- .
...
- Slots in the same task executor can have different resources. Ideally, to improve overall resource utility, we should allocate to a task a subset of resources that exactly matches its resource requirements. Since individual tasks may have different resource needs, the slots should also have resources.
- Dynamically create and destroy slots. Different jobs may need to partition the task executor’s resources into slots differently, depending on the particular resource requirements of tasks. Even for the same job, later tasks that trying to reuse resources released by previous finished tasks may prefer a different partition over the resources. Thus, we propose to partition a task executor’s resources dynamically, creating slots from available resources on demand, and destroying slots when they are released.
Task executors are launched with total resources but no predefined slots. When making allocation, instead of requesting an particular existing slot, the resource manager requests a slot with certain requested resources from the task executor. The task executor then create a new slot with the requested resources out of its available resources and offer the slot to the job master. As soon as the slot is released by the job master, it is destroyed and the its resources are returned back to the task executor as available resources.
Unknown Resource Requirements
Resource manager should always request slots from task executors with specified resource requirements. For slot requests with unknown resource requirements that it receives from job masters, it should allocate slots with default slot resource profiles from the task executors.
We introduce a config option defaultSlotFraction to configure what fraction of the task executor available resource a default slot should take. For compatibility, if defaultSlotFraction is not specified, we calculate it as 1 / numOfSlot, so that by default the task executor’s resources are partitioned in the same way as in the static slot model.
Given that in standalone clusters we may have different default slot resource for different task executors, we need task executors to register their default slot resource to the resource manager on registration. The default slot resource profile should only be calculated in either startup script (standalone) or resource manager (yarn / mesos / k8s), and passed into task executors as environment variables.
Protocol Changes
TaskExecutorGateway
Replace the requestSlot interface with a new requestResource interface.
...
requestSlot
...
requestResource
...
Parameters
...
- SlotID
- JobID
- AllocationID
- TargetAddress
- ResourceManagerID
...
- ResourceProfile
- JobID
- AllocationID
- TargetAddress
- ResourceManagerID
...
Return Value
...
Acknowledge
...
SlotID
...
A slot report that task executors send to the resource manager (in registration or heartbeats) now consists of two kinds of information.
- SlotStatus of allocated slots. A SlotStatus consists of the SlotID, ResourceProfile, JobID and AllocationID. Since slots are dynamically created and destroyed, there should not be any “free slot”. Therefore, the allocation id should never be null.
Available Resources of the task executor currently. Because of the asynchronous communication, it is possible for a resource manager to first make an allocation and then receive a slot report with outdated available resources. Later allocations based on that outdated available resources could fail due to insufficient available resources. In these cases, TaskExecutorGateway#requestResource will return null indicating allocation failure.
Compatibility, Deprecation, and Migration Plan
- This FLIP deprecates the configuration "taskmanager.numberOfTaskSlots", but stays should be compatible with itprevious versions.
Test Plan
- We need to update existing and add new integration tests dedicated to validate the new fine grained resource management behaviors.
- It is also expected that other regular integration and end-to-end tests should fail if this is broken.
Rejected Alternatives
An alternative for setting slot sharing groups in compiling is that, to set tasks with specified resource requirements into individual slot sharing groups (except for tasks in colocation groups), and tasks with unknown resource requirements in the same slot sharing group. It is rejected because it separates tasks from the same pipelined region into different slot sharing group, which may lead to a situation with resource deadlocks.
An alternative for setting relative managed memory quota for operators is to set it during the scheduling stage, with the knowledge of which tasks are actually scheduled into the same slot. It is rejected because even we make set the quota at scheduling, there is no guarantee that no tasks will be deployed into the same slot in the future, and dynamically updating the fractions requires operators’ memory usage to be revocable.