使用GPU的火花:如何强制每个执行器执行一个任务

Question

我在一个有N个从节点的集群上运行了Spark2.1.0。每个节点有16个核心(8个核心/cpu和2个cpu)和1个GPU。我想使用映射进程来启动GPU内核。因为每个节点只有一个GPU，所以我需要确保两个执行者不在同一个节点上(同时)尝试使用GPU，并且两个任务不会同时提交到同一个执行器。

如何强制每个节点拥有一个执行程序？

我尝试了以下几点：

-背景：spark.executor.cores 16 in $SPARK_HOME/conf/spark-defaults.conf

-背景：SPARK_WORKER_CORES = 16和SPARK_WORKER_INSTANCES = 1 in $SPARK_HOME/conf/spark-env.sh

和,

--直接在我的spark脚本中设置conf = SparkConf().set('spark.executor.cores', 16).set('spark.executor.instances', 6) (当我希望N=6用于调试时)。

这些选项按照需要在不同的节点上创建了6个执行程序，但是似乎每个任务都分配给了相同的执行器。

下面是我最近输出的一些片段(这让我相信它应该能像我想的那样工作)。

17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor added: app-20170217110910-0000/0 on worker-20170217110853-10.128.14.208-35771 (10.128.14.208:35771) with 16 cores
17/02/17 11:09:10 INFO StandaloneSchedulerBackend: Granted executor ID app-20170217110910-0000/0 on hostPort 10.128.14.208:35771 with 16 cores, 16.0 GB RAM 
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor added: app-20170217110910-0000/1 on worker-20170217110853-10.128.9.95-59294 (10.128.9.95:59294) with 16 cores
17/02/17 11:09:10 INFO StandaloneSchedulerBackend: Granted executor ID app-20170217110910-0000/1 on hostPort 10.128.9.95:59294 with 16 cores, 16.0 GB RAM 
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor added: app-20170217110910-0000/2 on worker-20170217110853-10.128.3.71-47507 (10.128.3.71:47507) with 16 cores
17/02/17 11:09:10 INFO StandaloneSchedulerBackend: Granted executor ID app-20170217110910-0000/2 on hostPort 10.128.3.71:47507 with 16 cores, 16.0 GB RAM 
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor added: app-20170217110910-0000/3 on worker-20170217110853-10.128.9.96-50800 (10.128.9.96:50800) with 16 cores
17/02/17 11:09:10 INFO StandaloneSchedulerBackend: Granted executor ID app-20170217110910-0000/3 on hostPort 10.128.9.96:50800 with 16 cores, 16.0 GB RAM 
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor added: app-20170217110910-0000/4 on worker-20170217110853-10.128.3.73-60194 (10.128.3.73:60194) with 16 cores
17/02/17 11:09:10 INFO StandaloneSchedulerBackend: Granted executor ID app-20170217110910-0000/4 on hostPort 10.128.3.73:60194 with 16 cores, 16.0 GB RAM 
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor added: app-20170217110910-0000/5 on worker-20170217110853-10.128.3.74-42793 (10.128.3.74:42793) with 16 cores
17/02/17 11:09:10 INFO StandaloneSchedulerBackend: Granted executor ID app-20170217110910-0000/5 on hostPort 10.128.3.74:42793 with 16 cores, 16.0 GB RAM 
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor updated: app-20170217110910-0000/1 is now RUNNING
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor updated: app-20170217110910-0000/3 is now RUNNING
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor updated: app-20170217110910-0000/4 is now RUNNING
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor updated: app-20170217110910-0000/2 is now RUNNING
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor updated: app-20170217110910-0000/0 is now RUNNING
17/02/17 11:09:10 INFO StandaloneAppClient$ClientEndpoint: Executor updated: app-20170217110910-0000/5 is now RUNNING
17/02/17 11:09:11 INFO StandaloneSchedulerBackend: SchedulerBackend is ready for scheduling beginning after reached minRegisteredResourcesRatio: 0.0 

My RDD has 6 partitions.

重要的是启动了6个执行者，每个执行者都有一个不同的IP地址，每个执行程序都有16个核心(与我所期望的完全相同)。短语My RDD has 6 partitions.是在重新划分我的RDD之后从我的代码中的一个打印语句(以确保每个执行器有一个分区)。

然后，这个发生了..。这6个任务中的每一个都被发送到同一个执行器！

17/02/17 11:09:12 INFO TaskSchedulerImpl: Adding task set 0.0 with 6 tasks
17/02/17 11:09:17 INFO CoarseGrainedSchedulerBackend$DriverEndpoint: Registered executor NettyRpcEndpointRef(null) (10.128.9.95:34059) with ID 1
17/02/17 11:09:17 INFO TaskSetManager: Starting task 0.0 in stage 0.0 (TID 0, 10.128.9.95, executor 1, partition 0, PROCESS_LOCAL, 6095 bytes)
17/02/17 11:09:17 INFO TaskSetManager: Starting task 1.0 in stage 0.0 (TID 1, 10.128.9.95, executor 1, partition 1, PROCESS_LOCAL, 6095 bytes)
17/02/17 11:09:17 INFO TaskSetManager: Starting task 2.0 in stage 0.0 (TID 2, 10.128.9.95, executor 1, partition 2, PROCESS_LOCAL, 6095 bytes)
17/02/17 11:09:17 INFO TaskSetManager: Starting task 3.0 in stage 0.0 (TID 3, 10.128.9.95, executor 1, partition 3, PROCESS_LOCAL, 6095 bytes)
17/02/17 11:09:17 INFO TaskSetManager: Starting task 4.0 in stage 0.0 (TID 4, 10.128.9.95, executor 1, partition 4, PROCESS_LOCAL, 6095 bytes)
17/02/17 11:09:17 INFO TaskSetManager: Starting task 5.0 in stage 0.0 (TID 5, 10.128.9.95, executor 1, partition 5, PROCESS_LOCAL, 6095 bytes)

为什么？和如何修复？问题是，此时，所有6个任务都在竞争同一个GPU，而GPU不能共享。

Answer 1

我试着在萨姆森·沙罗特的评论中提出这个建议，但这些建议似乎行不通。然而，我发现：http://spark.apache.org/docs/latest/configuration.html#scheduling，其中包括spark.task.cpus。如果我将其设置为16，而spark.executor.cores设置为16，那么我似乎得到了分配给每个执行器的一个任务。