Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Not A Bug
-
3.0.0
-
None
-
None
Description
I am running Spark in a local-cluster[2,1,1024] with one GPU per worker, task and executor, and two GPUs provided through a GPU discovery script. The same GPU is allocated to both executors.
Discovery script output:
{"name": "gpu", "addresses": ["0", "1"]}
Spark local cluster setup through spark-shell:
./spark-3.0.0-bin-hadoop2.7/bin/spark-shell --master "local-cluster[2,1,1024]" --conf spark.worker.resource.gpu.discoveryScript=/tmp/gpu.json --conf spark.worker.resource.gpu.amount=1 --conf spark.task.resource.gpu.amount=1 --conf spark.executor.resource.gpu.amount=1
Executor page of this cluster:
You can see that both executors have the same GPU allocated: [1]
Code run in the Spark shell:
scala> import org.apache.spark.TaskContext import org.apache.spark.TaskContext scala> def fn(it: Iterator[java.lang.Long]): Iterator[(String, (String, Array[String]))] = { TaskContext.get().resources().mapValues(v => (v.name, v.addresses)).iterator } fn: (it: Iterator[Long])Iterator[(String, (String, Array[String]))] scala> spark.range(0,2,1,2).mapPartitions(fn).collect res0: Array[(String, (String, Array[String]))] = Array((gpu,(gpu,Array(1))), (gpu,(gpu,Array(1))))
The result shows that each task got GPU 1. The executor page shows that each task has been run on different executors (see above screenshot).
The expected behaviour would have been to have GPU 0 assigned to one executor and GPU 1 to the other executor. Consequently, each partition / task should then see a different GPU.
With Spark 3.0.0-preview2 the allocation was as expected (identical code and Spark shell setup):
res0: Array[(String, (String, Array[String]))] = Array((gpu,(gpu,Array(0))), (gpu,(gpu,Array(1))))
Happy to contribute a patch if this is an accepted bug.
