Re: understanding failing my job, Giraph/Hadoop memory usage, under-utilized nodes, and moving forward
Sorry, should be *org.apache.giraph.utils.MemoryUtils.getRuntimeMemoryStats(), *I left out the giraph. On Mon, Sep 22, 2014 at 8:10 PM, Matthew Saltz sal...@gmail.com wrote: Hi Matthew, I answered a few of your questions in-line (unfortunately they might not help the larger problem, but hopefully it'll help a bit). Best, Matthew On Mon, Sep 22, 2014 at 5:50 PM, Matthew Cornell m...@matthewcornell.org wrote: Hi Folks, I've spent the last two months learning, installing, coding, and analyzing the performance of our Giraph app, and I'm able to run on small inputs on our tiny cluster (yay!) I am now stuck trying to figure out why larger inputs fail, why only some compute nodes are being used, and generally how to make sure I've configured hadoop and giraph to use all available CPUs and RAM. I feel that I'm this close, and I could really use some pointers. Below I share our app, configuration, results and log messages, some questions, and counter output for the successful run. My post here is long (I've broken it into sections delimited with ''), but I hope I've provided good enough information to get help on. I'm happy to add to it. Thanks! application Our application is a kind of path search where all nodes have a type and source database ID (e.g., movie 99), and searches are expressed as type paths, such as movie, acted_in, actor, which would start with movies and then find all actors in each movie, for all movies in the database. The program does a kind of filtering by keeping track of previously-processed initial IDs. Our database is a small movie one with 2K movies, 6K users (people who rate movies), and 80K ratings of movies by users. Though small, we've found this kind of search can result in a massive explosion of messages, as was well put by Rob Vesse ( http://mail-archives.apache.org/mod_mbox/giraph-user/201312.mbox/%3ccec4a409.2d7ad%25rve...@dotnetrdf.org%3E ): even with this relatively small graph you get a massive explosion of messages by the later super steps which exhausts memory (in my graph the theoretical maximum messages by the last super step was ~3 billion) job failure and error messages Currently I have a four-step path that completes in ~20 seconds (rates, movie, rates, user - counter output shown at bottom) but a five-step one (rates, movie, rates, user, rates) fails after a few minutes. I've looked carefully at the task logs, but I find it a little difficult to discern what the actual failure was. However, looking at system information (e.g., top and ganglia) during the run indicates hosts are running out of memory. There are no OutOfMemoryErrors in the logs, and only this one stsands out: ERROR org.apache.giraph.master.BspServiceMaster: superstepChosenWorkerAlive: Missing chosen worker Worker(hostname=compute-0-3.wright, MRtaskID=1, port=30001) on superstep 4 NB: So far I've been ignoring these other types of messages: FATAL org.apache.giraph.master.BspServiceMaster: getLastGoodCheckpoint: No last good checkpoints can be found, killing the job. java.io.FileNotFoundException: File _bsp/_checkpoints/job_201409191450_0003 does not exist. WARN org.apache.giraph.bsp.BspService: process: Unknown and unprocessed event (path=/_hadoopBsp/job_201409191450_0003/_applicationAttemptsDir/0/_superstepDir/2/_superstepFinished, type=NodeDeleted, state=SyncConnected) ERROR org.apache.giraph.worker.BspServiceWorker: unregisterHealth: Got failure, unregistering health on /_hadoopBsp/job_201409191450_0003/_applicationAttemptsDir/0/_superstepDir/4/_workerHealthyDir/compute-0-3.wright_1 on superstep 4 The counter statistics are minimal after the run fails, but during it I see something like this when refreshing the Job Tracker Web UI: Counters Map-Reduce Framework Physical memory (bytes) snapshot: ~28GB Counters Map-Reduce Framework Virtual memory (bytes) snapshot: ~27GB Counters Giraph Stats Sent messages: ~181M hadoop/giraph command hadoop jar $GIRAPH_HOME/giraph-ex.jar org.apache.giraph.GiraphRunner \ -Dgiraph.zkList=xx.xx.xx.edu:2181 \ -libjars ${LIBJARS} \ relpath.RelPathVertex \ -wc relpath.RelPathWorkerContext \ -mc relpath.RelPathMasterCompute \ -vif relpath.CausalityJsonAdjacencyListVertexInputFormat \ -vip $REL_PATH_INPUT \ -of relpath.RelPathVertexValueTextOutputFormat \ -op $REL_PATH_OUTPUT \ -ca RelPathVertex.path=$REL_PATH_PATH \ -w 8 cluster, versions, and configuration We have a five-node cluster with a head and four compute nodes. The head has 2 CPUs, 16 cores each, and 64 GB RAM. Each compute has 1 CPU, 4 cores each, and 16 GB RAM, making a total cluster of 128 GB of RAM and 48 cores. Hadoop: Cloudera CDH4 with a mapreduce service running the job tracker on the head node, and task trackers on all five nodes. Hadoop configuration (mapred-site.xml and CDH interface - sorry for the mix) - not sure I'm listing
Re: understanding failing my job, Giraph/Hadoop memory usage, under-utilized nodes, and moving forward
Hi Matthew, I answered a few of your questions in-line (unfortunately they might not help the larger problem, but hopefully it'll help a bit). Best, Matthew On Mon, Sep 22, 2014 at 5:50 PM, Matthew Cornell m...@matthewcornell.org wrote: Hi Folks, I've spent the last two months learning, installing, coding, and analyzing the performance of our Giraph app, and I'm able to run on small inputs on our tiny cluster (yay!) I am now stuck trying to figure out why larger inputs fail, why only some compute nodes are being used, and generally how to make sure I've configured hadoop and giraph to use all available CPUs and RAM. I feel that I'm this close, and I could really use some pointers. Below I share our app, configuration, results and log messages, some questions, and counter output for the successful run. My post here is long (I've broken it into sections delimited with ''), but I hope I've provided good enough information to get help on. I'm happy to add to it. Thanks! application Our application is a kind of path search where all nodes have a type and source database ID (e.g., movie 99), and searches are expressed as type paths, such as movie, acted_in, actor, which would start with movies and then find all actors in each movie, for all movies in the database. The program does a kind of filtering by keeping track of previously-processed initial IDs. Our database is a small movie one with 2K movies, 6K users (people who rate movies), and 80K ratings of movies by users. Though small, we've found this kind of search can result in a massive explosion of messages, as was well put by Rob Vesse ( http://mail-archives.apache.org/mod_mbox/giraph-user/201312.mbox/%3ccec4a409.2d7ad%25rve...@dotnetrdf.org%3E ): even with this relatively small graph you get a massive explosion of messages by the later super steps which exhausts memory (in my graph the theoretical maximum messages by the last super step was ~3 billion) job failure and error messages Currently I have a four-step path that completes in ~20 seconds (rates, movie, rates, user - counter output shown at bottom) but a five-step one (rates, movie, rates, user, rates) fails after a few minutes. I've looked carefully at the task logs, but I find it a little difficult to discern what the actual failure was. However, looking at system information (e.g., top and ganglia) during the run indicates hosts are running out of memory. There are no OutOfMemoryErrors in the logs, and only this one stsands out: ERROR org.apache.giraph.master.BspServiceMaster: superstepChosenWorkerAlive: Missing chosen worker Worker(hostname=compute-0-3.wright, MRtaskID=1, port=30001) on superstep 4 NB: So far I've been ignoring these other types of messages: FATAL org.apache.giraph.master.BspServiceMaster: getLastGoodCheckpoint: No last good checkpoints can be found, killing the job. java.io.FileNotFoundException: File _bsp/_checkpoints/job_201409191450_0003 does not exist. WARN org.apache.giraph.bsp.BspService: process: Unknown and unprocessed event (path=/_hadoopBsp/job_201409191450_0003/_applicationAttemptsDir/0/_superstepDir/2/_superstepFinished, type=NodeDeleted, state=SyncConnected) ERROR org.apache.giraph.worker.BspServiceWorker: unregisterHealth: Got failure, unregistering health on /_hadoopBsp/job_201409191450_0003/_applicationAttemptsDir/0/_superstepDir/4/_workerHealthyDir/compute-0-3.wright_1 on superstep 4 The counter statistics are minimal after the run fails, but during it I see something like this when refreshing the Job Tracker Web UI: Counters Map-Reduce Framework Physical memory (bytes) snapshot: ~28GB Counters Map-Reduce Framework Virtual memory (bytes) snapshot: ~27GB Counters Giraph Stats Sent messages: ~181M hadoop/giraph command hadoop jar $GIRAPH_HOME/giraph-ex.jar org.apache.giraph.GiraphRunner \ -Dgiraph.zkList=xx.xx.xx.edu:2181 \ -libjars ${LIBJARS} \ relpath.RelPathVertex \ -wc relpath.RelPathWorkerContext \ -mc relpath.RelPathMasterCompute \ -vif relpath.CausalityJsonAdjacencyListVertexInputFormat \ -vip $REL_PATH_INPUT \ -of relpath.RelPathVertexValueTextOutputFormat \ -op $REL_PATH_OUTPUT \ -ca RelPathVertex.path=$REL_PATH_PATH \ -w 8 cluster, versions, and configuration We have a five-node cluster with a head and four compute nodes. The head has 2 CPUs, 16 cores each, and 64 GB RAM. Each compute has 1 CPU, 4 cores each, and 16 GB RAM, making a total cluster of 128 GB of RAM and 48 cores. Hadoop: Cloudera CDH4 with a mapreduce service running the job tracker on the head node, and task trackers on all five nodes. Hadoop configuration (mapred-site.xml and CDH interface - sorry for the mix) - not sure I'm listing all of them of interest: mapreduce.job.counters.max: 120 mapred.output.compress: false mapred.reduce.tasks: 12 mapred.child.java.opts: -Xmx2147483648