...

...

Discussion thread: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-FLIP-53-Fine-Grained-Resource-Management-td31831.html

JIRA:

...

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

...

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.

Release notes: Since on-heap managed memory is removed in the final implementation of FLIP-49, fracManagedMemOnHeap and its calculation are also removed from the final implementation of this FLIP.

Slot Sharing

During the compiling stage, the StreamingJobGraphGenerator 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.

The StreamingJobGraphGenerator sets tasks of different pipelined regions into different slot sharing groups. In this way, when the StreamingJobGraphGenerator sets relative managed memory quota for operators, it will calculate the fractions only considering operators that might run at the same time. This improves resource utility utilisation for bounded batch jobs where usually not all tasks run concurrently.

...

To solve this problem, we need to put different connected components into the same slot sharing group for streaming jobs, while keep them in different slot sharing groups to avoid having large slots with tasks not necessarily scheduled together. We need a parameter scheduleAllSourcesTogether/allVerticesInSameSlotSharingGroupByDefault indicating whether to identify all the sources as in the same pipelined region (imagine a virtual source connected to all the real sources) or not, and passed it into StreamingJobGraphGenerator differently for streaming and batch jobs.

...

• 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.

Implementation Steps

Step 1. Introduce

...

option allVerticesInSameSlotSharingGroupByDefault in ExecutionConfig

• Introduce option allSourcesInSamePipelinedRegion in allVerticesInSameSlotSharingGroupByDefault in ExecutionConfig
• Set it to true by default
• Set it to false for SQL/Table API bounded batch jobs by the Blink planner

...

• Identify pipelined regions, with respect to allSourcesInSamePipelinedRegionto allVerticesInSameSlotSharingGroupByDefault
• Set slot sharing groups according to pipelined regions 
  • By default, each pipelined region should go into a separate slot sharing group
  • If the user sets operators in multiple pipelined regions into same slot sharing group, it should be respected

...

This step should not introduce any behavior changes.

Step 5. Operators use fractions to decide how

...

much managed memory to allocate

• Operators allocate memory segments with the amount returned by MemoryManager#computeNumberOfPages.
• Operators reserve memory with the amount returned by MemoryManager#computeMemorySize. 

...