Re: Flink Runtime Exception
I think Andra wrote that there is *no deadlock*. On Fri, Jun 19, 2015 at 3:18 PM Fabian Hueske fhue...@gmail.com http://mailto:fhue...@gmail.com wrote: Hi Andra, The system should never deadlock. There is a bug somewhere if that happens. Can you check if the program is really stuck? Cheers, Fabian 2015-06-19 15:08 GMT+02:00 Till Rohrmann trohrm...@apache.org: What does forever mean? Usually it's the case that you see a steep decline in performance once the system starts spilling data to disk because of the disk I/O bottleneck. The system always starts spilling to disk when it has no more memory left for its operations. So for example if you want to sort data which cannot be kept completely in memory, then the system has to employ external sorting. If you can give Flink more memory, then you can avoid this problem depending on the actual data size. You can observe disk I/O by using the `iotop` command for example. If you see Flink having a high I/O usage, then this might be an indicator for spilling. Cheers, Till On Fri, Jun 19, 2015 at 2:54 PM Andra Lungu lungu.an...@gmail.com wrote: Another problem that I encountered during the same set of experiments(sorry if I am asking too many questions, I am eager to get things fixed): - for the same configuration, a piece of code runs perfectly on 10GB of input, then for 38GB it runs forever (no deadlock). I believe that may occur because Flink spills information to disk every time it runs out of memory... Is this fixable by increasing the number of buffers? That's the last question for today, promise :) Thanks! On Fri, Jun 19, 2015 at 2:40 PM, Till Rohrmann trohrm...@apache.org wrote: Yes, it was an issue for the milestone release. On Fri, Jun 19, 2015 at 2:18 PM Andra Lungu lungu.an...@gmail.com wrote: Yes, so I am using flink-0.9.0-milestone-1. Was it a problem for this version? I'll just fetch the latest master if this is the case. On Fri, Jun 19, 2015 at 2:12 PM, Till Rohrmann trohrm...@apache.org wrote: Hi Andra, the problem seems to be that the deployment of some tasks takes longer than 100s. From the stack trace it looks as if you're not using the latest master. We had problems with previous version where the deployment call waited for the TM to completely download the user code jars. For large setups the BlobServer became a bottleneck and some of the deployment calls timed out. We updated the deployment logic so that the TM sends an immediate ACK backt to the JM when it receives a new task. Could you verify which version of Flink you're running and in case that it's not the latest master, could you please try to run your example with the latest code? Cheers, Till On Fri, Jun 19, 2015 at 1:42 PM Andra Lungu lungu.an...@gmail.com wrote: Hi everyone, I ran a job this morning on 30 wally nodes. DOP 224. Worked like a charm. Then, I ran a similar job, on the exact same configuration, on the same input data set. The only difference between them is that the second job computes the degrees per vertex and, for vertices with degree higher than a user-defined threshold, it does a bit of magic(roughly a bunch of coGroups). The problem is that, even before the extra functions get called, I get the following type of exception: 06/19/2015 12:06:43 CHAIN FlatMap (FlatMap at fromDataSet(Graph.java:171)) - Combine(Distinct at fromDataSet(Graph.java:171))(222/224) switched to FAILED java.lang.IllegalStateException: Update task on instance 29073fb0b0957198a2b67569b042d56b @ wally004 - 8 slots - URL: akka.tcp:// flink@130.149.249.14:44528/user/taskmanager failed due to: at org.apache.flink.runtime.executiongraph.Execution$5.onFailure(Execution.java:860) at akka.dispatch.OnFailure.internal(Future.scala:228) at akka.dispatch.OnFailure.internal(Future.scala:227) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:174) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:171) at scala.PartialFunction$class.applyOrElse(PartialFunction.scala:118) at scala.runtime.AbstractPartialFunction.applyOrElse(AbstractPartialFunction.scala:25) at scala.concurrent.Future$anonfun$onFailure$1.apply(Future.scala:136) at scala.concurrent.Future$anonfun$onFailure$1.apply(Future.scala:134) at
Re: Flink Runtime Exception
woops, sorry! Whenever I read the word deadlock I getting a bit nervous and distracted ;-) 2015-06-19 15:21 GMT+02:00 Till Rohrmann trohrm...@apache.org: I think Andra wrote that there is *no deadlock*. On Fri, Jun 19, 2015 at 3:18 PM Fabian Hueske fhue...@gmail.com http://mailto:fhue...@gmail.com wrote: Hi Andra, The system should never deadlock. There is a bug somewhere if that happens. Can you check if the program is really stuck? Cheers, Fabian 2015-06-19 15:08 GMT+02:00 Till Rohrmann trohrm...@apache.org: What does forever mean? Usually it's the case that you see a steep decline in performance once the system starts spilling data to disk because of the disk I/O bottleneck. The system always starts spilling to disk when it has no more memory left for its operations. So for example if you want to sort data which cannot be kept completely in memory, then the system has to employ external sorting. If you can give Flink more memory, then you can avoid this problem depending on the actual data size. You can observe disk I/O by using the `iotop` command for example. If you see Flink having a high I/O usage, then this might be an indicator for spilling. Cheers, Till On Fri, Jun 19, 2015 at 2:54 PM Andra Lungu lungu.an...@gmail.com wrote: Another problem that I encountered during the same set of experiments(sorry if I am asking too many questions, I am eager to get things fixed): - for the same configuration, a piece of code runs perfectly on 10GB of input, then for 38GB it runs forever (no deadlock). I believe that may occur because Flink spills information to disk every time it runs out of memory... Is this fixable by increasing the number of buffers? That's the last question for today, promise :) Thanks! On Fri, Jun 19, 2015 at 2:40 PM, Till Rohrmann trohrm...@apache.org wrote: Yes, it was an issue for the milestone release. On Fri, Jun 19, 2015 at 2:18 PM Andra Lungu lungu.an...@gmail.com wrote: Yes, so I am using flink-0.9.0-milestone-1. Was it a problem for this version? I'll just fetch the latest master if this is the case. On Fri, Jun 19, 2015 at 2:12 PM, Till Rohrmann trohrm...@apache.org wrote: Hi Andra, the problem seems to be that the deployment of some tasks takes longer than 100s. From the stack trace it looks as if you're not using the latest master. We had problems with previous version where the deployment call waited for the TM to completely download the user code jars. For large setups the BlobServer became a bottleneck and some of the deployment calls timed out. We updated the deployment logic so that the TM sends an immediate ACK backt to the JM when it receives a new task. Could you verify which version of Flink you're running and in case that it's not the latest master, could you please try to run your example with the latest code? Cheers, Till On Fri, Jun 19, 2015 at 1:42 PM Andra Lungu lungu.an...@gmail.com wrote: Hi everyone, I ran a job this morning on 30 wally nodes. DOP 224. Worked like a charm. Then, I ran a similar job, on the exact same configuration, on the same input data set. The only difference between them is that the second job computes the degrees per vertex and, for vertices with degree higher than a user-defined threshold, it does a bit of magic(roughly a bunch of coGroups). The problem is that, even before the extra functions get called, I get the following type of exception: 06/19/2015 12:06:43 CHAIN FlatMap (FlatMap at fromDataSet(Graph.java:171)) - Combine(Distinct at fromDataSet(Graph.java:171))(222/224) switched to FAILED java.lang.IllegalStateException: Update task on instance 29073fb0b0957198a2b67569b042d56b @ wally004 - 8 slots - URL: akka.tcp:// flink@130.149.249.14:44528/user/taskmanager failed due to: at org.apache.flink.runtime.executiongraph.Execution$5.onFailure(Execution.java:860) at akka.dispatch.OnFailure.internal(Future.scala:228) at akka.dispatch.OnFailure.internal(Future.scala:227) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:174) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:171) at scala.PartialFunction$class.applyOrElse(PartialFunction.scala:118) at
Re: Flink Runtime Exception
On 19 Jun 2015, at 14:53, Andra Lungu lungu.an...@gmail.com wrote: Another problem that I encountered during the same set of experiments(sorry if I am asking too many questions, I am eager to get things fixed): - for the same configuration, a piece of code runs perfectly on 10GB of input, then for 38GB it runs forever (no deadlock). I believe that may occur because Flink spills information to disk every time it runs out of memory... Is this fixable by increasing the number of buffers? If you are referring to the number of network buffers configuration key, that should be unrelated. If this really is the issue, you can increase the heap size for the task managers. Have you confirmed your suspicion as Till suggested via iotop? :) – Ufuk
Re: Flink Runtime Exception
Hi Andra, the problem seems to be that the deployment of some tasks takes longer than 100s. From the stack trace it looks as if you're not using the latest master. We had problems with previous version where the deployment call waited for the TM to completely download the user code jars. For large setups the BlobServer became a bottleneck and some of the deployment calls timed out. We updated the deployment logic so that the TM sends an immediate ACK backt to the JM when it receives a new task. Could you verify which version of Flink you're running and in case that it's not the latest master, could you please try to run your example with the latest code? Cheers, Till On Fri, Jun 19, 2015 at 1:42 PM Andra Lungu lungu.an...@gmail.com wrote: Hi everyone, I ran a job this morning on 30 wally nodes. DOP 224. Worked like a charm. Then, I ran a similar job, on the exact same configuration, on the same input data set. The only difference between them is that the second job computes the degrees per vertex and, for vertices with degree higher than a user-defined threshold, it does a bit of magic(roughly a bunch of coGroups). The problem is that, even before the extra functions get called, I get the following type of exception: 06/19/2015 12:06:43 CHAIN FlatMap (FlatMap at fromDataSet(Graph.java:171)) - Combine(Distinct at fromDataSet(Graph.java:171))(222/224) switched to FAILED java.lang.IllegalStateException: Update task on instance 29073fb0b0957198a2b67569b042d56b @ wally004 - 8 slots - URL: akka.tcp:// flink@130.149.249.14:44528/user/taskmanager failed due to: at org.apache.flink.runtime.executiongraph.Execution$5.onFailure(Execution.java:860) at akka.dispatch.OnFailure.internal(Future.scala:228) at akka.dispatch.OnFailure.internal(Future.scala:227) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:174) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:171) at scala.PartialFunction$class.applyOrElse(PartialFunction.scala:118) at scala.runtime.AbstractPartialFunction.applyOrElse(AbstractPartialFunction.scala:25) at scala.concurrent.Future$$anonfun$onFailure$1.apply(Future.scala:136) at scala.concurrent.Future$$anonfun$onFailure$1.apply(Future.scala:134) at scala.concurrent.impl.CallbackRunnable.run(Promise.scala:32) at scala.concurrent.impl.ExecutionContextImpl$$anon$3.exec(ExecutionContextImpl.scala:107) at scala.concurrent.forkjoin.ForkJoinTask.doExec(ForkJoinTask.java:260) at scala.concurrent.forkjoin.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:1339) at scala.concurrent.forkjoin.ForkJoinPool.runWorker(ForkJoinPool.java:1979) at scala.concurrent.forkjoin.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:107) Caused by: akka.pattern.AskTimeoutException: Ask timed out on [Actor[akka.tcp://flink@130.149.249.14:44528/user/taskmanager#82700874]] after [10 ms] at akka.pattern.PromiseActorRef$$anonfun$1.apply$mcV$sp(AskSupport.scala:333) at akka.actor.Scheduler$$anon$7.run(Scheduler.scala:117) at scala.concurrent.Future$InternalCallbackExecutor$.scala$concurrent$Future$InternalCallbackExecutor$$unbatchedExecute(Future.scala:694) at scala.concurrent.Future$InternalCallbackExecutor$.execute(Future.scala:691) at akka.actor.LightArrayRevolverScheduler$TaskHolder.executeTask(Scheduler.scala:467) at akka.actor.LightArrayRevolverScheduler$$anon$8.executeBucket$1(Scheduler.scala:419) at akka.actor.LightArrayRevolverScheduler$$anon$8.nextTick(Scheduler.scala:423) at akka.actor.LightArrayRevolverScheduler$$anon$8.run(Scheduler.scala:375) at java.lang.Thread.run(Thread.java:722) At first I thought, okay maybe wally004 is down; then I ssh'd into it. Works fine. The full output can be found here: https://gist.github.com/andralungu/d222b75cb33aea57955d Does anyone have any idea about what may have triggered this? :( Thanks! Andra
Re: Flink Runtime Exception
Yes, so I am using flink-0.9.0-milestone-1. Was it a problem for this version? I'll just fetch the latest master if this is the case. On Fri, Jun 19, 2015 at 2:12 PM, Till Rohrmann trohrm...@apache.org wrote: Hi Andra, the problem seems to be that the deployment of some tasks takes longer than 100s. From the stack trace it looks as if you're not using the latest master. We had problems with previous version where the deployment call waited for the TM to completely download the user code jars. For large setups the BlobServer became a bottleneck and some of the deployment calls timed out. We updated the deployment logic so that the TM sends an immediate ACK backt to the JM when it receives a new task. Could you verify which version of Flink you're running and in case that it's not the latest master, could you please try to run your example with the latest code? Cheers, Till On Fri, Jun 19, 2015 at 1:42 PM Andra Lungu lungu.an...@gmail.com wrote: Hi everyone, I ran a job this morning on 30 wally nodes. DOP 224. Worked like a charm. Then, I ran a similar job, on the exact same configuration, on the same input data set. The only difference between them is that the second job computes the degrees per vertex and, for vertices with degree higher than a user-defined threshold, it does a bit of magic(roughly a bunch of coGroups). The problem is that, even before the extra functions get called, I get the following type of exception: 06/19/2015 12:06:43 CHAIN FlatMap (FlatMap at fromDataSet(Graph.java:171)) - Combine(Distinct at fromDataSet(Graph.java:171))(222/224) switched to FAILED java.lang.IllegalStateException: Update task on instance 29073fb0b0957198a2b67569b042d56b @ wally004 - 8 slots - URL: akka.tcp:// flink@130.149.249.14:44528/user/taskmanager failed due to: at org.apache.flink.runtime.executiongraph.Execution$5.onFailure(Execution.java:860) at akka.dispatch.OnFailure.internal(Future.scala:228) at akka.dispatch.OnFailure.internal(Future.scala:227) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:174) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:171) at scala.PartialFunction$class.applyOrElse(PartialFunction.scala:118) at scala.runtime.AbstractPartialFunction.applyOrElse(AbstractPartialFunction.scala:25) at scala.concurrent.Future$$anonfun$onFailure$1.apply(Future.scala:136) at scala.concurrent.Future$$anonfun$onFailure$1.apply(Future.scala:134) at scala.concurrent.impl.CallbackRunnable.run(Promise.scala:32) at scala.concurrent.impl.ExecutionContextImpl$$anon$3.exec(ExecutionContextImpl.scala:107) at scala.concurrent.forkjoin.ForkJoinTask.doExec(ForkJoinTask.java:260) at scala.concurrent.forkjoin.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:1339) at scala.concurrent.forkjoin.ForkJoinPool.runWorker(ForkJoinPool.java:1979) at scala.concurrent.forkjoin.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:107) Caused by: akka.pattern.AskTimeoutException: Ask timed out on [Actor[akka.tcp://flink@130.149.249.14:44528/user/taskmanager#82700874]] after [10 ms] at akka.pattern.PromiseActorRef$$anonfun$1.apply$mcV$sp(AskSupport.scala:333) at akka.actor.Scheduler$$anon$7.run(Scheduler.scala:117) at scala.concurrent.Future$InternalCallbackExecutor$.scala$concurrent$Future$InternalCallbackExecutor$$unbatchedExecute(Future.scala:694) at scala.concurrent.Future$InternalCallbackExecutor$.execute(Future.scala:691) at akka.actor.LightArrayRevolverScheduler$TaskHolder.executeTask(Scheduler.scala:467) at akka.actor.LightArrayRevolverScheduler$$anon$8.executeBucket$1(Scheduler.scala:419) at akka.actor.LightArrayRevolverScheduler$$anon$8.nextTick(Scheduler.scala:423) at akka.actor.LightArrayRevolverScheduler$$anon$8.run(Scheduler.scala:375) at java.lang.Thread.run(Thread.java:722) At first I thought, okay maybe wally004 is down; then I ssh'd into it. Works fine. The full output can be found here: https://gist.github.com/andralungu/d222b75cb33aea57955d Does anyone have any idea about what may have triggered this? :( Thanks! Andra
Re: Flink Runtime Exception
Yes, it was an issue for the milestone release. On Fri, Jun 19, 2015 at 2:18 PM Andra Lungu lungu.an...@gmail.com wrote: Yes, so I am using flink-0.9.0-milestone-1. Was it a problem for this version? I'll just fetch the latest master if this is the case. On Fri, Jun 19, 2015 at 2:12 PM, Till Rohrmann trohrm...@apache.org wrote: Hi Andra, the problem seems to be that the deployment of some tasks takes longer than 100s. From the stack trace it looks as if you're not using the latest master. We had problems with previous version where the deployment call waited for the TM to completely download the user code jars. For large setups the BlobServer became a bottleneck and some of the deployment calls timed out. We updated the deployment logic so that the TM sends an immediate ACK backt to the JM when it receives a new task. Could you verify which version of Flink you're running and in case that it's not the latest master, could you please try to run your example with the latest code? Cheers, Till On Fri, Jun 19, 2015 at 1:42 PM Andra Lungu lungu.an...@gmail.com wrote: Hi everyone, I ran a job this morning on 30 wally nodes. DOP 224. Worked like a charm. Then, I ran a similar job, on the exact same configuration, on the same input data set. The only difference between them is that the second job computes the degrees per vertex and, for vertices with degree higher than a user-defined threshold, it does a bit of magic(roughly a bunch of coGroups). The problem is that, even before the extra functions get called, I get the following type of exception: 06/19/2015 12:06:43 CHAIN FlatMap (FlatMap at fromDataSet(Graph.java:171)) - Combine(Distinct at fromDataSet(Graph.java:171))(222/224) switched to FAILED java.lang.IllegalStateException: Update task on instance 29073fb0b0957198a2b67569b042d56b @ wally004 - 8 slots - URL: akka.tcp:// flink@130.149.249.14:44528/user/taskmanager failed due to: at org.apache.flink.runtime.executiongraph.Execution$5.onFailure(Execution.java:860) at akka.dispatch.OnFailure.internal(Future.scala:228) at akka.dispatch.OnFailure.internal(Future.scala:227) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:174) at akka.dispatch.japi$CallbackBridge.apply(Future.scala:171) at scala.PartialFunction$class.applyOrElse(PartialFunction.scala:118) at scala.runtime.AbstractPartialFunction.applyOrElse(AbstractPartialFunction.scala:25) at scala.concurrent.Future$$anonfun$onFailure$1.apply(Future.scala:136) at scala.concurrent.Future$$anonfun$onFailure$1.apply(Future.scala:134) at scala.concurrent.impl.CallbackRunnable.run(Promise.scala:32) at scala.concurrent.impl.ExecutionContextImpl$$anon$3.exec(ExecutionContextImpl.scala:107) at scala.concurrent.forkjoin.ForkJoinTask.doExec(ForkJoinTask.java:260) at scala.concurrent.forkjoin.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:1339) at scala.concurrent.forkjoin.ForkJoinPool.runWorker(ForkJoinPool.java:1979) at scala.concurrent.forkjoin.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:107) Caused by: akka.pattern.AskTimeoutException: Ask timed out on [Actor[akka.tcp://flink@130.149.249.14:44528/user/taskmanager#82700874 ]] after [10 ms] at akka.pattern.PromiseActorRef$$anonfun$1.apply$mcV$sp(AskSupport.scala:333) at akka.actor.Scheduler$$anon$7.run(Scheduler.scala:117) at scala.concurrent.Future$InternalCallbackExecutor$.scala$concurrent$Future$InternalCallbackExecutor$$unbatchedExecute(Future.scala:694) at scala.concurrent.Future$InternalCallbackExecutor$.execute(Future.scala:691) at akka.actor.LightArrayRevolverScheduler$TaskHolder.executeTask(Scheduler.scala:467) at akka.actor.LightArrayRevolverScheduler$$anon$8.executeBucket$1(Scheduler.scala:419) at akka.actor.LightArrayRevolverScheduler$$anon$8.nextTick(Scheduler.scala:423) at akka.actor.LightArrayRevolverScheduler$$anon$8.run(Scheduler.scala:375) at java.lang.Thread.run(Thread.java:722) At first I thought, okay maybe wally004 is down; then I ssh'd into it. Works fine. The full output can be found here: https://gist.github.com/andralungu/d222b75cb33aea57955d Does anyone have any idea about what may have triggered this? :( Thanks! Andra
Re: Flink Runtime Exception
Hi Andra, I would try increasing the memory per task manager, i.e. on a machine with 8 CPUs and 16GBs of memory, instead of spawning 8 TMs with 2GB each, I would try to spawn 2 TMs of 8GBs each. This might help with the spilling problem (in case that the CPU is not your bottleneck, this might even speed up the computations by avoiding spilling) and get you unstuck. Cheers, Asterios On Fri, Jun 19, 2015 at 4:16 PM, Ufuk Celebi u...@apache.org wrote: On 19 Jun 2015, at 14:53, Andra Lungu lungu.an...@gmail.com wrote: Another problem that I encountered during the same set of experiments(sorry if I am asking too many questions, I am eager to get things fixed): - for the same configuration, a piece of code runs perfectly on 10GB of input, then for 38GB it runs forever (no deadlock). I believe that may occur because Flink spills information to disk every time it runs out of memory... Is this fixable by increasing the number of buffers? If you are referring to the number of network buffers configuration key, that should be unrelated. If this really is the issue, you can increase the heap size for the task managers. Have you confirmed your suspicion as Till suggested via iotop? :) – Ufuk
