Re: How many task managers to launch for a job?

Posted by black chase on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/How-many-task-managers-to-launch-for-a-job-tp27983p27999.html

Yes true. I am trying to figure out how the TaskManagers are distributed across physical machines by Mesos and YARN. Maybe I shoud start a new thread for help.

Thank you Song

Best,

Pengcheng

On Fri, May 24, 2019 at 10:52 AM Xintong Song <[hidden email]> wrote:

As far as I know, Flink does not have any requirements on how the TaskManagers are distributed across physical machines. So I think it really depends on the scheduling policy of the Mesos cluster. I'm not an expert on Mesos, so correct me if I was wrong.

On Fri, May 24, 2019 at 4:18 PM black chase <[hidden email]> wrote:

Hi Song,

Thank you for the clarification.

Now I know TaskManagers are automatically allocated. Yet, I am still not very clear how the TMs are allocated.

I'm guessing the allocation process would be:

On the job side, I have a job with each operator parallelism=5. Since one TaskManager has one slot. It means this job would need 5 TaskManagers becase one pipeline would need one slot according to the task scheduling policy.

On the Mesos side, let's say currently there are no available TaskManagers. Then Mesos would spawn 5 new TaskManagers for this job. Is it?

If what I am guessing is right, then let's say we have 4 physical computing nodes for the Flink TaskManagers. How would Mesos place the 5 new TaskManagers into the 4 physical computing nodes? Is it juts a Round-Robin fashion?

Kind regards

Chase

On Fri, May 24, 2019 at 4:10 AM Xintong Song <[hidden email]> wrote:

Hi black,

If you are running Flink on Yarn or Mesos, Flink will automatically allocate resource and launch new TaskManagers as needed.

If you are using Flink standalone mode, then the easiest way is to enable slot sharing and set all the vertices into the same group (which is by default). In that way, the total slots (or number of TaskManagers if you config on slot for each TaskManager) needed for running the job would be the maximum parallelism of the job graph vertices. Further information on slot sharing could be found here.

On Thu, May 23, 2019 at 11:49 PM black chase <[hidden email]> wrote:

Hi,

I am redesigning the scheduler of the JobManager to place tasks of a job across TaskManagers accroding to a scheduling policy.

I am reading the Flip-6 proposal and found that the common case is "one TaskManager launchs one slot", and "one Flink cluster serves one job". But I did not find how many TaskManagers to launch in a computing node. Is there any common practice for this ?

--

--

--