Github user susanxhuynh commented on a diff in the pull request:
https://github.com/apache/spark/pull/21033#discussion_r182283035
--- Diff:
resource-managers/mesos/src/main/scala/org/apache/spark/scheduler/cluster/mesos/MesosCoarseGrainedSchedulerBackend.scala
---
@@ -495,9 +500,8 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
launchTasks = true
val taskId = newMesosTaskId()
val offerCPUs = getResource(resources, "cpus").toInt
- val taskGPUs = Math.min(
- Math.max(0, maxGpus - totalGpusAcquired),
getResource(resources, "gpus").toInt)
-
+ val offerGPUs = getResource(resources, "gpus").toInt
+ var taskGPUs = executorGpus
--- End diff --
Ah, good point, I missed that earlier. @yanji84 Why are we changing the
default behavior when `spark.mesos.executor.gpus` is not specified? Previously,
if `spark.mesos.gpus.max` was set (without setting
`spark.mesos.executor.gpus`), GPUs were allocated greedily. This aligns with
the CPU behavior when `spark.executor.cores` is not specified. GPUs could be
handled the same way.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
