mllib performance on mesos cluster
Hi, I have a program similar to the BinaryClassifier example that I am running using my data (which is fairly small). I run this for 100 iterations. I observed the following performance: Standalone mode cluster with 10 nodes (with Spark 1.0.2): 5 minutes Standalone mode cluster with 10 nodes (with Spark 1.1.0): 8.9 minutes Mesos cluster with 10 nodes (with Spark 1.1.0): 17 minutes 1) According to the documentation, Spark 1.1.0 has better performance. So I would like to understand why the runtime is longer on Spark 1.1.0. 2) Why is the performance on mesos significantly higher than in standalone mode? I just wanted to find out if any one else has observed poor performance for Mllib based programs on mesos cluster. I looked through the application detail logs and found that some of the scheduler delay values were higher on mesos compared to standalone mode (40 ms vs. 10 ms). Is the mesos scheduler slower? thanks -- View this message in context: http://apache-spark-user-list.1001560.n3.nabble.com/mllib-performance-on-mesos-cluster-tp14692.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: mllib performance on mesos cluster
1) MLlib 1.1 should be faster than 1.0 in general. What's the size of your dataset? Is the RDD evenly distributed across nodes? You can check the storage tab of the Spark WebUI. 2) I don't have much experience with mesos deployment. Someone else may be able to answer your question. -Xiangrui On Fri, Sep 19, 2014 at 12:17 PM, SK wrote: > Hi, > > I have a program similar to the BinaryClassifier example that I am running > using my data (which is fairly small). I run this for 100 iterations. I > observed the following performance: > > Standalone mode cluster with 10 nodes (with Spark 1.0.2): 5 minutes > Standalone mode cluster with 10 nodes (with Spark 1.1.0): 8.9 minutes > Mesos cluster with 10 nodes (with Spark 1.1.0): 17 minutes > > 1) According to the documentation, Spark 1.1.0 has better performance. So I > would like to understand why the runtime is longer on Spark 1.1.0. > > 2) Why is the performance on mesos significantly higher than in standalone > mode? I just wanted to find out if any one else has observed poor > performance for Mllib based programs on mesos cluster. I looked through the > application detail logs and found that some of the scheduler delay values > were higher on mesos compared to standalone mode (40 ms vs. 10 ms). Is the > mesos scheduler slower? > > thanks > > > > -- > View this message in context: > http://apache-spark-user-list.1001560.n3.nabble.com/mllib-performance-on-mesos-cluster-tp14692.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 > - To unsubscribe, e-mail: user-unsubscr...@spark.apache.org For additional commands, e-mail: user-h...@spark.apache.org
Re: mllib performance on mesos cluster
Setting spark.mesos.coarse=true helped reduce the time on the mesos cluster from 17 min to around 6 min. The scheduler delay per task reduced from 40 ms to around 10 ms. thanks On Mon, Sep 22, 2014 at 12:36 PM, Xiangrui Meng wrote: > 1) MLlib 1.1 should be faster than 1.0 in general. What's the size of > your dataset? Is the RDD evenly distributed across nodes? You can > check the storage tab of the Spark WebUI. > > 2) I don't have much experience with mesos deployment. Someone else > may be able to answer your question. > > -Xiangrui > > On Fri, Sep 19, 2014 at 12:17 PM, SK wrote: > > Hi, > > > > I have a program similar to the BinaryClassifier example that I am > running > > using my data (which is fairly small). I run this for 100 iterations. I > > observed the following performance: > > > > Standalone mode cluster with 10 nodes (with Spark 1.0.2): 5 minutes > > Standalone mode cluster with 10 nodes (with Spark 1.1.0): 8.9 minutes > > Mesos cluster with 10 nodes (with Spark 1.1.0): 17 minutes > > > > 1) According to the documentation, Spark 1.1.0 has better performance. > So I > > would like to understand why the runtime is longer on Spark 1.1.0. > > > > 2) Why is the performance on mesos significantly higher than in > standalone > > mode? I just wanted to find out if any one else has observed poor > > performance for Mllib based programs on mesos cluster. I looked through > the > > application detail logs and found that some of the scheduler delay values > > were higher on mesos compared to standalone mode (40 ms vs. 10 ms). Is > the > > mesos scheduler slower? > > > > thanks > > > > > > > > -- > > View this message in context: > http://apache-spark-user-list.1001560.n3.nabble.com/mllib-performance-on-mesos-cluster-tp14692.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 > > >