[ https://issues.apache.org/jira/browse/SPARK-3100?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14100664#comment-14100664 ]
Ravindra Pesala edited comment on SPARK-3100 at 9/30/14 7:47 AM: ----------------------------------------------------------------- It seems there is an issue in synchronization of task allocation and registering of executors. As per the above log I observed that tasks has been allocated with single registered executor. After these tasks are allocated to registered executor,remaining executors are started registering. So this may be synchronization issue. I have added sleep in my driver for 5 seconds then it started working properly. {code} val sc = new SparkContext("spark://master:7077", "SampleSpark", "/opt/spark-1.0.0-rc3/",ar) *Thread.sleep(5000)* val rdd = new SampleRDD(sc,new KeyValImpl()); rdd.collect; {code} Now the tasks are assigned to all the nodes {code} INFO 18-08 19:44:36,301 - Registered executor: Actor[akka.tcp://sparkExecutor@master:32457/user/Executor#-966982652] with ID 0 INFO 18-08 19:44:36,505 - Registering block manager master:59964 with 294.4 MB RAM INFO 18-08 19:44:36,578 - Registered executor: Actor[akka.tcp://sparkExecutor@slave1:11653/user/Executor#909834981] with ID 1 INFO 18-08 19:44:36,591 - Registered executor: Actor[akka.tcp://sparkExecutor@slave2:59220/user/Executor#301495226] with ID 3 INFO 18-08 19:44:36,643 - Registered executor: Actor[akka.tcp://sparkExecutor@slave3:11232/user/Executor#-1118376183] with ID 2 INFO 18-08 19:44:36,804 - Registering block manager slave1:14596 with 294.4 MB RAM INFO 18-08 19:44:36,809 - Registering block manager slave2:10418 with 294.4 MB RAM INFO 18-08 19:44:36,871 - Registering block manager slave3:45973 with 294.4 MB RAM INFO 18-08 19:44:39,507 - Starting job: collect at SampleRDD.scala:142 INFO 18-08 19:44:39,520 - Got job 0 (collect at SampleRDD.scala:142) with 4 output partitions (allowLocal=false) INFO 18-08 19:44:39,521 - Final stage: Stage 0(collect at SampleRDD.scala:142) INFO 18-08 19:44:39,521 - Parents of final stage: List() INFO 18-08 19:44:39,526 - Missing parents: List() INFO 18-08 19:44:39,532 - Submitting Stage 0 (SampleRDD[0] at RDD at SampleRDD.scala:28), which has no missing parents INFO 18-08 19:44:39,537 - Host Name : master INFO 18-08 19:44:39,539 - Host Name : slave1 INFO 18-08 19:44:39,539 - Host Name : slave2 INFO 18-08 19:44:39,540 - Host Name : slave3 INFO 18-08 19:44:39,563 - Submitting 4 missing tasks from Stage 0 (SampleRDD[0] at RDD at SampleRDD.scala:28) INFO 18-08 19:44:39,564 - Adding task set 0.0 with 4 tasks INFO 18-08 19:44:39,579 - Starting task 0.0:2 as TID 0 on executor 3: *slave2 (NODE_LOCAL)* INFO 18-08 19:44:39,583 - Serialized task 0.0:2 as 1261 bytes in 2 ms INFO 18-08 19:44:39,585 - Starting task 0.0:0 as TID 1 on executor 0: *master (NODE_LOCAL)* INFO 18-08 19:44:39,585 - Serialized task 0.0:0 as 1261 bytes in 0 ms INFO 18-08 19:44:39,586 - Starting task 0.0:1 as TID 2 on executor 1: *slave1 (NODE_LOCAL)* INFO 18-08 19:44:39,586 - Serialized task 0.0:1 as 1261 bytes in 0 ms INFO 18-08 19:44:39,587 - Starting task 0.0:3 as TID 3 on executor 2: *slave3 (NODE_LOCAL)* INFO 18-08 19:44:39,587 - Serialized task 0.0:3 as 1261 bytes in 0 ms {code} Is it expected behavior? Please comment on it. was (Author: ravipesala): It seems there is an issue in synchronization of task allocation and registering of executors. As per the above log I observed that task allocations are done with single registered executor. After these tasks are started ,remaining executors are started registering. So this is synchronization issue. I have added sleep in my driver for 5 seconds then it started working properly. val sc = new SparkContext("spark://master:7077", "SampleSpark", "/opt/spark-1.0.0-rc3/",ar) *Thread.sleep(5000)* val rdd = new SampleRDD(sc,new KeyValImpl()); rdd.collect; INFO 18-08 19:44:36,301 - Registered executor: Actor[akka.tcp://sparkExecutor@master:32457/user/Executor#-966982652] with ID 0 INFO 18-08 19:44:36,505 - Registering block manager master:59964 with 294.4 MB RAM INFO 18-08 19:44:36,578 - Registered executor: Actor[akka.tcp://sparkExecutor@slave1:11653/user/Executor#909834981] with ID 1 INFO 18-08 19:44:36,591 - Registered executor: Actor[akka.tcp://sparkExecutor@slave2:59220/user/Executor#301495226] with ID 3 INFO 18-08 19:44:36,643 - Registered executor: Actor[akka.tcp://sparkExecutor@slave3:11232/user/Executor#-1118376183] with ID 2 INFO 18-08 19:44:36,804 - Registering block manager slave1:14596 with 294.4 MB RAM INFO 18-08 19:44:36,809 - Registering block manager slave2:10418 with 294.4 MB RAM INFO 18-08 19:44:36,871 - Registering block manager slave3:45973 with 294.4 MB RAM INFO 18-08 19:44:39,507 - Starting job: collect at SampleRDD.scala:142 INFO 18-08 19:44:39,520 - Got job 0 (collect at SampleRDD.scala:142) with 4 output partitions (allowLocal=false) INFO 18-08 19:44:39,521 - Final stage: Stage 0(collect at SampleRDD.scala:142) INFO 18-08 19:44:39,521 - Parents of final stage: List() INFO 18-08 19:44:39,526 - Missing parents: List() INFO 18-08 19:44:39,532 - Submitting Stage 0 (SampleRDD[0] at RDD at SampleRDD.scala:28), which has no missing parents INFO 18-08 19:44:39,537 - Host Name : master INFO 18-08 19:44:39,539 - Host Name : slave1 INFO 18-08 19:44:39,539 - Host Name : slave2 INFO 18-08 19:44:39,540 - Host Name : slave3 INFO 18-08 19:44:39,563 - Submitting 4 missing tasks from Stage 0 (SampleRDD[0] at RDD at SampleRDD.scala:28) INFO 18-08 19:44:39,564 - Adding task set 0.0 with 4 tasks INFO 18-08 19:44:39,579 - Starting task 0.0:2 as TID 0 on executor 3: *slave2 (NODE_LOCAL)* INFO 18-08 19:44:39,583 - Serialized task 0.0:2 as 1261 bytes in 2 ms INFO 18-08 19:44:39,585 - Starting task 0.0:0 as TID 1 on executor 0: *master (NODE_LOCAL)* INFO 18-08 19:44:39,585 - Serialized task 0.0:0 as 1261 bytes in 0 ms INFO 18-08 19:44:39,586 - Starting task 0.0:1 as TID 2 on executor 1: *slave1 (NODE_LOCAL)* INFO 18-08 19:44:39,586 - Serialized task 0.0:1 as 1261 bytes in 0 ms INFO 18-08 19:44:39,587 - Starting task 0.0:3 as TID 3 on executor 2: *slave3 (NODE_LOCAL)* INFO 18-08 19:44:39,587 - Serialized task 0.0:3 as 1261 bytes in 0 ms > Spark RDD partitions are not running in the workers as per locality > information given by each partition. > -------------------------------------------------------------------------------------------------------- > > Key: SPARK-3100 > URL: https://issues.apache.org/jira/browse/SPARK-3100 > Project: Spark > Issue Type: Bug > Components: Spark Core > Affects Versions: 1.0.0 > Environment: Running in Spark Standalone Cluster > Reporter: Ravindra Pesala > > I created a simple custom RDD (SampleRDD.scala)and created 4 splits for 4 > workers. > When I run this RDD in Spark standalone cluster with 4 workers(even master > machine has one worker), it runs all partitions in one node only even though > I have given locality preferences in my SampleRDD program. > *Sample Code* > {code} > class SamplePartition(rddId: Int, val idx: Int,val tableSplit:Seq[String]) > extends Partition { > override def hashCode(): Int = 41 * (41 + rddId) + idx > override val index: Int = idx > } > class SampleRDD[K,V]( > sc : SparkContext,keyClass: KeyVal[K,V]) > extends RDD[(K,V)](sc, Nil) > with Logging { > override def getPartitions: Array[Partition] = { > val hosts = Array("master","slave1","slave2","slave3") > val result = new Array[Partition](4) > for (i <- 0 until result.length) > { > result(i) = new SamplePartition(id, i, Array(hosts(i))) > } > result > } > > > override def compute(theSplit: Partition, context: TaskContext) = { > val iter = new Iterator[(K,V)] { > val split = theSplit.asInstanceOf[SamplePartition] > logInfo("Executed task for the split" + split.tableSplit) > > // Register an on-task-completion callback to close the input stream. > context.addOnCompleteCallback(() => close()) > var havePair = false > var finished = false > override def hasNext: Boolean = { > if (!finished && !havePair) > { > finished = !false > havePair = !finished > } > !finished > } > override def next(): (K,V) = { > if (!hasNext) { > throw new java.util.NoSuchElementException("End of stream") > } > havePair = false > val key = new Key() > val value = new Value() > keyClass.getKey(key, value) > } > private def close() { > try { > // reader.close() > } catch { > case e: Exception => logWarning("Exception in > RecordReader.close()", e) > } > } > } > iter > } > > override def getPreferredLocations(split: Partition): Seq[String] = { > val theSplit = split.asInstanceOf[SamplePartition] > val s = theSplit.tableSplit.filter(_ != "localhost") > logInfo("Host Name : "+s(0)) > s > } > } > trait KeyVal[K,V] extends Serializable { > def getKey(key : Key,value : Value) : (K,V) > } > class KeyValImpl extends KeyVal[Key,Value] { > override def getKey(key : Key,value : Value) = (key,value) > } > case class Key() > case class Value() > object SampleRDD { > def main(args: Array[String]) : Unit= { > val d = SparkContext.jarOfClass(this.getClass) > val ar = new Array[String](d.size) > var i = 0 > d.foreach{ > p=> ar(i)=p; > i = i+1 > } > val sc = new SparkContext("spark://master:7077", "SampleSpark", > "/opt/spark-1.0.0-rc3/",ar) > val rdd = new SampleRDD(sc,new KeyValImpl()); > rdd.collect; > } > } > {code} > Following is the log it shows. > {code} > INFO 18-08 16:38:33,382 - Executor updated: app-20140818163833-0005/0 is now > RUNNING > INFO 18-08 16:38:33,382 - Executor updated: app-20140818163833-0005/2 is now > RUNNING > INFO 18-08 16:38:33,383 - Executor updated: app-20140818163833-0005/1 is now > RUNNING > INFO 18-08 16:38:33,385 - Executor updated: app-20140818163833-0005/3 is now > RUNNING > INFO 18-08 16:38:34,976 - Registered executor: Actor > akka.tcp://sparkExecutor@master:47563/user/Executor#-398354094 with ID 0 > INFO 18-08 16:38:34,984 - Starting task 0.0:0 as TID 0 on executor 0: master > (PROCESS_LOCAL) > INFO 18-08 16:38:34,989 - Serialized task 0.0:0 as 1261 bytes in 3 ms > INFO 18-08 16:38:34,992 - Starting task 0.0:1 as TID 1 on executor 0: master > (PROCESS_LOCAL) > INFO 18-08 16:38:34,993 - Serialized task 0.0:1 as 1261 bytes in 0 ms > INFO 18-08 16:38:34,993 - Starting task 0.0:2 as TID 2 on executor 0: master > (PROCESS_LOCAL)* > INFO 18-08 16:38:34,993 - Serialized task 0.0:2 as 1261 bytes in 0 ms > INFO 18-08 16:38:34,994 - Starting task 0.0:3 as TID 3 on executor 0: master > (PROCESS_LOCAL) > INFO 18-08 16:38:34,994 - Serialized task 0.0:3 as 1261 bytes in 0 ms > INFO 18-08 16:38:35,174 - Registering block manager master:42125 with 294.4 > MB RAM > INFO 18-08 16:38:35,296 - Registered executor: Actor > akka.tcp://sparkExecutor@slave1:31726/user/Executor#492173410 with ID 2 > INFO 18-08 16:38:35,302 - Registered executor: Actor > akka.tcp://sparkExecutor@slave2:25769/user/Executor#1762839887 with ID 1 > INFO 18-08 16:38:35,317 - Registered executor: Actor > akka.tcp://sparkExecutor@slave3:51032/user/Executor#981476000 with ID 3 > {code} > Here all the tasks are assigned to master only, even I though I have > mentioned the locality preferences -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org