Run more than nr-of-cores concurrent tasks.

[Erik]

We are using the terraform databricks provier, which is starting a cluster and checking every mount (since there is no mount rest API!). Each mount takes 20 seconds to check, and 99.9% of that time is idle waiting, and it starts a job per mount. If we could run many (more than nr of cores) jobs concurrently we should be able to make it faster, but I cant find how to do this. I have tried setting

`spark.executor.instances` to 2*cores, but it seems to be ignored.

So, is it possible to set databricks to use more spark executors than nr of cores?

[Erik]

Hey @Kaniz Fatma​ , it does not seem like the community have an answer to this. Maybe you have access to some Databricks engineers who know the answer?

[jose_gonzalez]

Hi @Erik Parmann​ ,

Does this old post helps link

Also, where did you added the Spark configuration for "spark.executor.instances"? this should be set at the cluster level setting.

[Erik]

Hi @Jose Gonzalez​, thanks for the suggestion. But that link asks how to *limit* the nr of executors, so each gets more memory. I want to do the opposite, I want *more* executors per core (or make each executor execute many parallell tasks). The default value for `spark.task.cpus` is `1`, and it does not seem to accept a value like 0.1, then it refuses to start up.

I set the cluster level settings under "Advanced options", below I attached a screenshot of how I tried editing the spark.task.cpus setting:

[Erik]

@Jose Gonzalez​ @Kaniz Fatma​  : Since there is no more answers I am starting to belive that maybe it is not possible to get databricks to use more spark executors than nr of cores. Can you verify this for me?

[jose_gonzalez]

hi @Erik Parmann​ ,

It is possible to do, but you might need to also enable dynamic allocation at the cluster level to be able to make sure your settings are apply at cluster creation . You can find more details here. As best practice, we do not recommend to change this configurations because it might create other issues. We recommend to use the default options we provided.

[Erik]

Thanks for the reply! I understand that in generall the default options are good, but in this exact usecase (many tiny operations which are each 99.99999% IO bound) it is really suboptimal, and it really make the databricks-with-IAC experience a bit cumbersome.

I tried with the following settings in the "Spark Config" section:

spark.dynamicAllocation.enabled true
spark.dynamicAllocation.shuffleTracking.enabled true
spark.shuffle.service.enabled true
spark.dynamicAllocation.initialExecutors 8
spark.dynamicAllocation.minExecutors 8
spark.scheduler.mode FIFO

 

But on a 4-core machine I am still only able to get 1 executor (as seen in the "Spark Cluster UI"-tab) executing up to 4 tasks in parallell. I tried with "High concurrency" cluster and "Standard". Are you actually able to get many executors running by changing "spark.dynamicAllocation.enabled" and "spark.dynamicAllocation.minExecutors" ?