Major Description
While executing the SimpleShortestPaths example with Giraph 1.1.0-SNAPSHOT compiled for Hadoop Yarn 2.2.0, I've noticed that the application would never stop even after recognizing that all supersteps had completed and the output had been written to the output directory.
Looking at the logs, I found that the BspServiceMaster is stuck at the while loop at the end of cleanrUpZooKeeper() (BspServiceMaster.java:1729):
2013-12-08 03:51:21,698 INFO [org.apache.giraph.master.MasterThread] master.MasterThread (MasterThread.java:run(121)) - masterThread: Coordination of superstep 3 took 0.433 seconds ended with state ALL_SUPERSTEPS_DONE and is now on superstep 4
2013-12-08 03:51:21,699 INFO [org.apache.giraph.master.MasterThread] master.BspServiceMaster (BspServiceMaster.java:setJobState(261)) - setJobState: {"_stateKey":"FINISHED","_applicationAttemptKey":-1,"_superstepKey":-1} on superstep 4
2013-12-08 03:51:21,753 INFO [org.apache.giraph.master.MasterThread] master.BspServiceMaster (BspServiceMaster.java:cleanup(1836)) - cleanup: Notifying master its okay to cleanup with /_hadoopBsp/giraph_yarn_application_1386468390622_0005/_cleanedUpDir/0_master
2013-12-08 03:51:21,790 INFO [org.apache.giraph.master.MasterThread] master.BspServiceMaster (BspServiceMaster.java:cleanUpZooKeeper(1711)) - cleanUpZooKeeper: Node /_hadoopBsp/giraph_yarn_application_1386468390622_0005/_cleanedUpDir already exists, no need to create.
2013-12-08 03:51:21,792 INFO [org.apache.giraph.master.MasterThread] bsp.BspInputFormat (BspInputFormat.java:getMaxTasks(64)) - getMaxTasks: Max workers = 1, split master/worker = true, is YARN-only job = true, total max tasks = 1
2013-12-08 03:51:21,792 INFO [org.apache.giraph.master.MasterThread] master.BspServiceMaster (BspServiceMaster.java:cleanUpZooKeeper(1735)) - cleanUpZooKeeper: Got 2 of 1 desired children from /_hadoopBsp/giraph_yarn_application_1386468390622_0005/_cleanedUpDir
2013-12-08 03:51:21,793 INFO [org.apache.giraph.master.MasterThread] master.BspServiceMaster (BspServiceMaster.java:cleanUpZooKeeper(1744)) - cleanedUpZooKeeper: Waiting for the children of /_hadoopBsp/giraph_yarn_application_1386468390622_0005/_cleanedUpDir to change since only got 2 nodes.
As the last 2 entries show, instead of registering just 1 task ending, it registers 2 and thus it misses the condition on line 1740.
One solution would be to change the == in line 1740 to a >=. However, the actual issue seems to reside with the BspInputFormat.getMaxTasks() (BspInputFormat.java:51). This function assumes that in a pure yarn execution the total number of tasks will be equal to the maximum number of workers. However, based on GiraphApplicationMaster:167, this is not the case. An extra Master task is launched in addition to all the Worker tasks. BspInputFormat.getMaxTasks() should then return maxWorkers + 1 in the case of a pure yarn execution.
Compilation:
mvn -Phadoop_yarn -Dhadoop.version=2.2.0 -DskipTests compile
Execution command:
$HADOOP_PREFIX/bin/hadoop jar ~/Projects/giraph/giraph-examples/target/giraph-examples-1.1.0-SNAPSHOT-for-hadoop-2.2.0-jar-with-dependencies.jar org.apache.giraph.GiraphRunner org.apache.giraph.examples.SimpleShortestPathsComputation -vif org.apache.giraph.io.formats.JsonLongDoubleFloatDoubleVertexInputFormat -vip giraph/input/tiny_graph.txt -vof org.apache.giraph.io.formats.IdWithValueTextOutputFormat -op giraph/output/shortestpahts -w 1 -ca giraph.zkList=localhost:2181 -yj giraph-examples-1.1.0-SNAPSHOT-for-hadoop-2.2.0-jar-with-dependencies.jar