Task distribution uneven

[HongyuLi-ms]

Thanks for filing an issue with us! Below are some guidelines when filing an issue. In general, the more detail the better!

Bug Report:

• Actual behavior
  some executors could get many tasks which the number is more than cores

for example ,there are 8 cores per executor. 120 accepts 10 tasks, but 61 accepts 5 tasks. That will cause the job spent more time .

• Expected behavior
  the number of executors accept tasks won't exceed their cores

• test result of mine
  I found driver assign tasks evenly if I set preFetchCount=2. also I tested kafka-spark-connector and it's has no problem. distribution evenly.
  So I wonder maybe that was caused by the cache, preFetchCount causes the memory usage is not as we expected. so driver won't assign tasks anymore.

• Spark version
  3.1.2

• spark-eventhubs artifactId and version
  com.microsoft.azure:azure-eventhubs-spark_2.12:2.3.21