Re: Any ideas why a few tasks would stall
Hi Steve et al., It is possible that there's just a lot of skew in your data, in which case repartitioning is a good idea. Depending on how large your input data is and how much skew you have, you may want to repartition to a larger number of partitions. By the way you can just call rdd.repartition(1000); this is the same as rdd.coalesce(1000, forceShuffle = true). Note that repartitioning is only a good idea if your straggler task is taking a long time. Otherwise, it can be quite expensive since it requires a full shuffle. Another possibility is that you might just have bad nodes in your cluster. To mitigate stragglers, you can try enabling speculative execution through spark.speculation to true. This attempts to re-run any task that takes a long time to complete on a different node in parallel. -Andrew 2014-12-04 11:43 GMT-08:00 akhandeshi ami.khande...@gmail.com: This did not work for me. that is, rdd.coalesce(200, forceShuffle) . Does anyone have ideas on how to distribute your data evenly and co-locate partitions of interest? -- View this message in context: http://apache-spark-user-list.1001560.n3.nabble.com/Any-ideas-why-a-few-tasks-would-stall-tp20207p20387.html Sent from the Apache Spark User List mailing list archive at Nabble.com. - To unsubscribe, e-mail: user-unsubscr...@spark.apache.org For additional commands, e-mail: user-h...@spark.apache.org
Re: Any ideas why a few tasks would stall
I ran into something similar before. 19/20 partitions would complete very quickly, and 1 would take the bulk of time and shuffle reads writes. This was because the majority of partitions were empty, and 1 had all the data. Perhaps something similar is going on here - I would suggest taking a look at how much data each partition contains and try to achieve a roughly even distribution for best performance. In particular, if the RDDs are PairRDDs, partitions are assigned based on the hash of the key, so an even distribution of values among keys is required for even split of data across partitions. On December 2, 2014 at 4:15:25 PM, Steve Lewis (lordjoe2...@gmail.com) wrote: 1) I can go there but none of the links are clickable 2) when I see something like 116/120 partitions succeeded in the stages ui in the storage ui I see NOTE RDD 27 has 116 partitions cached - 4 not and those are exactly the number of machines which will not complete Also RDD 27 does not show up in the Stages UI RDD NameStorage Level Cached Partitions Fraction Cached Size in Memory Size in Tachyon Size on Disk 2 Memory Deserialized 1x Replicated 1 100%11.8 MB 0.0 B 0.0 B 14 Memory Deserialized 1x Replicated 1 100%122.7 MB 0.0 B 0.0 B 7 Memory Deserialized 1x Replicated 120 100%151.1 MB 0.0 B 0.0 B 1 Memory Deserialized 1x Replicated 1 100%65.6 MB 0.0 B 0.0 B 10 Memory Deserialized 1x Replicated 24 100%160.6 MB 0.0 B 0.0 B 27 Memory Deserialized 1x Replicated 116 97% On Tue, Dec 2, 2014 at 3:43 PM, Sameer Farooqui same...@databricks.com wrote: Have you tried taking thread dumps via the UI? There is a link to do so on the Executors' page (typically under http://driver IP:4040/exectuors. By visualizing the thread call stack of the executors with slow running tasks, you can see exactly what code is executing at an instant in time. If you sample the executor several times in a short time period, you can identify 'hot spots' or expensive sections in the user code. On Tue, Dec 2, 2014 at 3:03 PM, Steve Lewis lordjoe2...@gmail.com wrote: I am working on a problem which will eventually involve many millions of function calls. A have a small sample with several thousand calls working but when I try to scale up the amount of data things stall. I use 120 partitions and 116 finish in very little time. The remaining 4 seem to do all the work and stall after a fixed number (about 1000) calls and even after hours make no more progress. This is my first large and complex job with spark and I would like any insight on how to debug the issue or even better why it might exist. The cluster has 15 machines and I am setting executor memory at 16G. Also what other questions are relevant to solving the issue -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com
Re: Any ideas why a few tasks would stall
Good point, Ankit. Steve - You can click on the link for '27' in the first column to get a break down of how much data is in each of those 116 cached partitions. But really, you want to also understand how much data is in the 4 non-cached partitions, as they may be huge. One thing you can try doing is .repartition() on the RDD with something like 100 partitions and then cache this new RDD. See if that spreads the load between the partitions more evenly. Let us know how it goes. On Thu, Dec 4, 2014 at 12:16 AM, Ankit Soni ankitso...@gmail.com wrote: I ran into something similar before. 19/20 partitions would complete very quickly, and 1 would take the bulk of time and shuffle reads writes. This was because the majority of partitions were empty, and 1 had all the data. Perhaps something similar is going on here - I would suggest taking a look at how much data each partition contains and try to achieve a roughly even distribution for best performance. In particular, if the RDDs are PairRDDs, partitions are assigned based on the hash of the key, so an even distribution of values among keys is required for even split of data across partitions. On December 2, 2014 at 4:15:25 PM, Steve Lewis (lordjoe2...@gmail.com) wrote: 1) I can go there but none of the links are clickable 2) when I see something like 116/120 partitions succeeded in the stages ui in the storage ui I see NOTE RDD 27 has 116 partitions cached - 4 not and those are exactly the number of machines which will not complete Also RDD 27 does not show up in the Stages UI RDD Name Storage Level Cached Partitions Fraction Cached Size in Memory Size in Tachyon Size on Disk 2 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=2 Memory Deserialized 1x Replicated 1 100% 11.8 MB 0.0 B 0.0 B 14 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=14 Memory Deserialized 1x Replicated 1 100% 122.7 MB 0.0 B 0.0 B 7 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=7 Memory Deserialized 1x Replicated 120 100% 151.1 MB 0.0 B 0.0 B 1 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=1 Memory Deserialized 1x Replicated 1 100% 65.6 MB 0.0 B 0.0 B 10 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=10 Memory Deserialized 1x Replicated 24 100% 160.6 MB 0.0 B 0.0 B 27 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=27 Memory Deserialized 1x Replicated 116 97% On Tue, Dec 2, 2014 at 3:43 PM, Sameer Farooqui same...@databricks.com wrote: Have you tried taking thread dumps via the UI? There is a link to do so on the Executors' page (typically under http://driver IP:4040/exectuors. By visualizing the thread call stack of the executors with slow running tasks, you can see exactly what code is executing at an instant in time. If you sample the executor several times in a short time period, you can identify 'hot spots' or expensive sections in the user code. On Tue, Dec 2, 2014 at 3:03 PM, Steve Lewis lordjoe2...@gmail.com wrote: I am working on a problem which will eventually involve many millions of function calls. A have a small sample with several thousand calls working but when I try to scale up the amount of data things stall. I use 120 partitions and 116 finish in very little time. The remaining 4 seem to do all the work and stall after a fixed number (about 1000) calls and even after hours make no more progress. This is my first large and complex job with spark and I would like any insight on how to debug the issue or even better why it might exist. The cluster has 15 machines and I am setting executor memory at 16G. Also what other questions are relevant to solving the issue -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com
Re: Any ideas why a few tasks would stall
Thanks - I found the same thing - calling boolean forceShuffle = true; myRDD = myRDD.coalesce(120,forceShuffle ); worked - there were 120 partitions but forcing a shuffle distributes the work I believe there is a bug in my code causing memory to accumulate as partitions grow in size. With a job ofer ten times larger I ran into other issues raising the number of partitions to 10,000 - namely too many open files On Thu, Dec 4, 2014 at 8:32 AM, Sameer Farooqui same...@databricks.com wrote: Good point, Ankit. Steve - You can click on the link for '27' in the first column to get a break down of how much data is in each of those 116 cached partitions. But really, you want to also understand how much data is in the 4 non-cached partitions, as they may be huge. One thing you can try doing is .repartition() on the RDD with something like 100 partitions and then cache this new RDD. See if that spreads the load between the partitions more evenly. Let us know how it goes. On Thu, Dec 4, 2014 at 12:16 AM, Ankit Soni ankitso...@gmail.com wrote: I ran into something similar before. 19/20 partitions would complete very quickly, and 1 would take the bulk of time and shuffle reads writes. This was because the majority of partitions were empty, and 1 had all the data. Perhaps something similar is going on here - I would suggest taking a look at how much data each partition contains and try to achieve a roughly even distribution for best performance. In particular, if the RDDs are PairRDDs, partitions are assigned based on the hash of the key, so an even distribution of values among keys is required for even split of data across partitions. On December 2, 2014 at 4:15:25 PM, Steve Lewis (lordjoe2...@gmail.com) wrote: 1) I can go there but none of the links are clickable 2) when I see something like 116/120 partitions succeeded in the stages ui in the storage ui I see NOTE RDD 27 has 116 partitions cached - 4 not and those are exactly the number of machines which will not complete Also RDD 27 does not show up in the Stages UI RDD Name Storage Level Cached Partitions Fraction Cached Size in Memory Size in Tachyon Size on Disk 2 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=2 Memory Deserialized 1x Replicated 1 100% 11.8 MB 0.0 B 0.0 B 14 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=14 Memory Deserialized 1x Replicated 1 100% 122.7 MB 0.0 B 0.0 B 7 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=7 Memory Deserialized 1x Replicated 120 100% 151.1 MB 0.0 B 0.0 B 1 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=1 Memory Deserialized 1x Replicated 1 100% 65.6 MB 0.0 B 0.0 B 10 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=10 Memory Deserialized 1x Replicated 24 100% 160.6 MB 0.0 B 0.0 B 27 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=27 Memory Deserialized 1x Replicated 116 97% On Tue, Dec 2, 2014 at 3:43 PM, Sameer Farooqui same...@databricks.com wrote: Have you tried taking thread dumps via the UI? There is a link to do so on the Executors' page (typically under http://driver IP:4040/exectuors. By visualizing the thread call stack of the executors with slow running tasks, you can see exactly what code is executing at an instant in time. If you sample the executor several times in a short time period, you can identify 'hot spots' or expensive sections in the user code. On Tue, Dec 2, 2014 at 3:03 PM, Steve Lewis lordjoe2...@gmail.com wrote: I am working on a problem which will eventually involve many millions of function calls. A have a small sample with several thousand calls working but when I try to scale up the amount of data things stall. I use 120 partitions and 116 finish in very little time. The remaining 4 seem to do all the work and stall after a fixed number (about 1000) calls and even after hours make no more progress. This is my first large and complex job with spark and I would like any insight on how to debug the issue or even better why it might exist. The cluster has 15 machines and I am setting executor memory at 16G. Also what other questions are relevant to solving the issue -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com
Re: Any ideas why a few tasks would stall
This did not work for me. that is, rdd.coalesce(200, forceShuffle) . Does anyone have ideas on how to distribute your data evenly and co-locate partitions of interest? -- View this message in context: http://apache-spark-user-list.1001560.n3.nabble.com/Any-ideas-why-a-few-tasks-would-stall-tp20207p20387.html Sent from the Apache Spark User List mailing list archive at Nabble.com. - To unsubscribe, e-mail: user-unsubscr...@spark.apache.org For additional commands, e-mail: user-h...@spark.apache.org
Re: Any ideas why a few tasks would stall
Have you tried taking thread dumps via the UI? There is a link to do so on the Executors' page (typically under http://driver IP:4040/exectuors. By visualizing the thread call stack of the executors with slow running tasks, you can see exactly what code is executing at an instant in time. If you sample the executor several times in a short time period, you can identify 'hot spots' or expensive sections in the user code. On Tue, Dec 2, 2014 at 3:03 PM, Steve Lewis lordjoe2...@gmail.com wrote: I am working on a problem which will eventually involve many millions of function calls. A have a small sample with several thousand calls working but when I try to scale up the amount of data things stall. I use 120 partitions and 116 finish in very little time. The remaining 4 seem to do all the work and stall after a fixed number (about 1000) calls and even after hours make no more progress. This is my first large and complex job with spark and I would like any insight on how to debug the issue or even better why it might exist. The cluster has 15 machines and I am setting executor memory at 16G. Also what other questions are relevant to solving the issue
Re: Any ideas why a few tasks would stall
1) I can go there but none of the links are clickable 2) when I see something like 116/120 partitions succeeded in the stages ui in the storage ui I see NOTE RDD 27 has 116 partitions cached - 4 not and those are exactly the number of machines which will not complete Also RDD 27 does not show up in the Stages UI RDD NameStorage LevelCached PartitionsFraction CachedSize in MemorySize in TachyonSize on Disk2 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=2Memory Deserialized 1x Replicated1100%11.8 MB0.0 B0.0 B14 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=14Memory Deserialized 1x Replicated1100%122.7 MB0.0 B0.0 B7 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=7Memory Deserialized 1x Replicated120100%151.1 MB0.0 B0.0 B1 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=1Memory Deserialized 1x Replicated1100%65.6 MB0.0 B0.0 B10 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=10Memory Deserialized 1x Replicated24100%160.6 MB0.0 B0.0 B27 http://hwlogin.labs.uninett.no:4040/storage/rdd?id=27Memory Deserialized 1x Replicated11697% On Tue, Dec 2, 2014 at 3:43 PM, Sameer Farooqui same...@databricks.com wrote: Have you tried taking thread dumps via the UI? There is a link to do so on the Executors' page (typically under http://driver IP:4040/exectuors. By visualizing the thread call stack of the executors with slow running tasks, you can see exactly what code is executing at an instant in time. If you sample the executor several times in a short time period, you can identify 'hot spots' or expensive sections in the user code. On Tue, Dec 2, 2014 at 3:03 PM, Steve Lewis lordjoe2...@gmail.com wrote: I am working on a problem which will eventually involve many millions of function calls. A have a small sample with several thousand calls working but when I try to scale up the amount of data things stall. I use 120 partitions and 116 finish in very little time. The remaining 4 seem to do all the work and stall after a fixed number (about 1000) calls and even after hours make no more progress. This is my first large and complex job with spark and I would like any insight on how to debug the issue or even better why it might exist. The cluster has 15 machines and I am setting executor memory at 16G. Also what other questions are relevant to solving the issue -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com