Flink evenly distributes the required slots across all Task Managers

Question

I have two different tasks in my flink job Task1 & Task2. T1 needs 10 Parrelisms and T2 needs 2 parrallism, my cluster has two Taskmanager(TM) and each of my task managers have 8 slots. Task1 runs with default task groups and for Task 2 I m setting separate slot-sharing group

Since I had requirements to distribute my talks sots across TM evenly, I added a cluster. evenly-spread-out-slots: true and deploy my application but evenly distributed is not working. Task manager 1 runs 8 (2 from Task2 and 6 from Task1) and Taskmanager 2 runs 4.

My requirement is each task manager should run 6 (5 for Task1 and 1 for Task2)

FOr testing, whether evenly-spread-out-slots: true work runs or not I commented out Task1 and submitted Task2, with this I can see Task2 runs on both TM1 and TM2 with 1 slot each.

Is there any setting to make sure evenly distributions happen for both the task groups?

score 2 · Answer 1 · answered Aug 22 '22 at 08:14

I had a similar problem with even slots distribution and ended up utilizing the Fine-Grained Resource Management.

So, in your case, let's assume that each task manager has 6Gb RAM and 6 cores, for example. Then, in order to have exactly 5 slots for task-1 slot-sharing group per-tm and exactly 1 slot for task-2 slot sharing group per-tm, I used mutual exclusive requirements:

final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();

SlotSharingGroup ssgTask1 = SlotSharingGroup.newBuilder("task-1")
  .setCpuCores(1.1)
  .setTaskHeapMemoryMB(10)
  .build();

SlotSharingGroup ssgTask2 = SlotSharingGroup.newBuilder("task-2")
  .setCpuCores(0.1)
  .setTaskHeapMemoryMB(5900)
  .build();

With this specification, task-2 slot sharing group requires slot with 5900MB heap, which almost occupies all the task manager capacity and therefore each task manager is able to allocate at most 1 task-2 slot.

Similarly, task-1 slot sharing group requires 1.1 cpu cores, so with 6 cores in total per-tm, each tm is able to allocate at most 5 task-1 slots.

With this slot sharing groups configuration I always get even slot sharing group slots distribution across all task managers.

That's a clever solution. – David Anderson Aug 22 '22 at 16:32 — David Anderson, Aug 22 '22 at 16:32

David Anderson · Answer 2 · 2022-08-21T17:04:04.440

0

With evenly-spread-out-slots: false

the scheduler will take slots from TM1 until none remain, and then move on to TM2

With evenly-spread-out-slots: true

the scheduler will take slots from the least used TM when there aren’t any other preferences
sources don't have preferences
the rest of the topology will follow, with consumers preferring to be co-located with their producers

I not sure it's possible to achieve what you're trying to do. But have you tried this without putting task 2 into a separate slot-sharing group?

edited Aug 21 '22 at 17:04

answered Aug 21 '22 at 16:55

David Anderson

39,434
4
33
60

Yes, Task1 get distributes evenly but both slots of task2 is running on the same taskmanager – scoder Aug 21 '22 at 17:53

Flink evenly distributes the required slots across all Task Managers

2 Answers2