How to speed PySpark to match Scala/Java performance
Hi, In my company, we've been trying to use PySpark to run ETLs on our data. Alas, it turned out to be terribly slow compared to Java or Scala API (which we ended up using to meet performance criteria). To be more quantitative, let's consider simple case: I've generated test file (848MB): /seq 1 1 > /tmp/test/ and tried to run simple computation on it, which includes three steps: read -> multiply each row by 2 -> take max Code in python: /sc.textFile("/tmp/test").map(lambda x: x * 2).max()/ Code in scala: /sc.textFile("/tmp/test").map(x => x * 2).max()/ Here are the results of this simple benchmark: CPython - 59s PyPy - 26s Scala version - 7s I didn't dig into what exactly contributes to execution times of CPython / PyPy, but it seems that serialization / deserialization, when sending data to the worker may be the issue. I know some guys already have been asking about using Jython (http://apache-spark-developers-list.1001551.n3.nabble.com/Jython-importing-pyspark-td8654.html#a8658, http://apache-spark-developers-list.1001551.n3.nabble.com/PySpark-Driver-from-Jython-td7142.html), but it seems, that no one have really done this with Spark. It looks like performance gain from using jython can be huge - you wouldn't need to spawn PythonWorkers, all the code would be just executed inside SparkExecutor JVM, using python code compiled to java bytecode. Do you think that's possible to achieve? Do you see any obvious obstacles? Of course, jython doesn't have C extensions, but if one doesn't need them, then it should fit here nicely. I'm willing to try to marry Spark with Jython and see how it goes. What do you think about this? -- View this message in context: http://apache-spark-developers-list.1001551.n3.nabble.com/How-to-speed-PySpark-to-match-Scala-Java-performance-tp10356.html Sent from the Apache Spark Developers List mailing list archive at Nabble.com. - To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org For additional commands, e-mail: dev-h...@spark.apache.org
Re: How to speed PySpark to match Scala/Java performance
Hey, Without having Python as fast as Scala/Java, I think it's impossible to similar performance in PySpark as in Scala/Java. Jython is also much slower than Scala/Java. With Jython, we can avoid the cost of manage multiple process and RPC, we may still need to do the data conversion between Java and Python. Given that fact that Jython is not widely used in production, it may introduce more troubles than the performance gain. Spark jobs can be easily speed up by scaling out (by adding more resources). I think the most advantage of PySpark is that it let you do fast prototype. Once you got your ETL finalized, it's not that hard to translate your pure Python jobs into Scala to reduce the cost(it's optional). Now days, engineer time is much more expensive than CPU time, I think we should be more focus on the former. That's my 2 cents. Davies On Thu, Jan 29, 2015 at 12:45 PM, rtshadow wrote: > Hi, > > In my company, we've been trying to use PySpark to run ETLs on our data. > Alas, it turned out to be terribly slow compared to Java or Scala API (which > we ended up using to meet performance criteria). > > To be more quantitative, let's consider simple case: > I've generated test file (848MB): /seq 1 1 > /tmp/test/ > > and tried to run simple computation on it, which includes three steps: read > -> multiply each row by 2 -> take max > Code in python: /sc.textFile("/tmp/test").map(lambda x: x * 2).max()/ > Code in scala: /sc.textFile("/tmp/test").map(x => x * 2).max()/ > > Here are the results of this simple benchmark: > CPython - 59s > PyPy - 26s > Scala version - 7s > > I didn't dig into what exactly contributes to execution times of CPython / > PyPy, but it seems that serialization / deserialization, when sending data > to the worker may be the issue. > I know some guys already have been asking about using Jython > (http://apache-spark-developers-list.1001551.n3.nabble.com/Jython-importing-pyspark-td8654.html#a8658, > http://apache-spark-developers-list.1001551.n3.nabble.com/PySpark-Driver-from-Jython-td7142.html), > but it seems, that no one have really done this with Spark. > It looks like performance gain from using jython can be huge - you wouldn't > need to spawn PythonWorkers, all the code would be just executed inside > SparkExecutor JVM, using python code compiled to java bytecode. Do you think > that's possible to achieve? Do you see any obvious obstacles? Of course, > jython doesn't have C extensions, but if one doesn't need them, then it > should fit here nicely. > > I'm willing to try to marry Spark with Jython and see how it goes. > > What do you think about this? > > > > > > -- > View this message in context: > http://apache-spark-developers-list.1001551.n3.nabble.com/How-to-speed-PySpark-to-match-Scala-Java-performance-tp10356.html > Sent from the Apache Spark Developers List mailing list archive at Nabble.com. > > - > To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org > For additional commands, e-mail: dev-h...@spark.apache.org > - To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org For additional commands, e-mail: dev-h...@spark.apache.org
Re: How to speed PySpark to match Scala/Java performance
Once the data frame API is released for 1.3, you can write your thing in Python and get the same performance. It can't express everything, but for basic things like projection, filter, join, aggregate and simple numeric computation, it should work pretty well. On Thu, Jan 29, 2015 at 12:45 PM, rtshadow wrote: > Hi, > > In my company, we've been trying to use PySpark to run ETLs on our data. > Alas, it turned out to be terribly slow compared to Java or Scala API > (which > we ended up using to meet performance criteria). > > To be more quantitative, let's consider simple case: > I've generated test file (848MB): /seq 1 1 > /tmp/test/ > > and tried to run simple computation on it, which includes three steps: read > -> multiply each row by 2 -> take max > Code in python: /sc.textFile("/tmp/test").map(lambda x: x * 2).max()/ > Code in scala: /sc.textFile("/tmp/test").map(x => x * 2).max()/ > > Here are the results of this simple benchmark: > CPython - 59s > PyPy - 26s > Scala version - 7s > > I didn't dig into what exactly contributes to execution times of CPython / > PyPy, but it seems that serialization / deserialization, when sending data > to the worker may be the issue. > I know some guys already have been asking about using Jython > ( > http://apache-spark-developers-list.1001551.n3.nabble.com/Jython-importing-pyspark-td8654.html#a8658 > , > > http://apache-spark-developers-list.1001551.n3.nabble.com/PySpark-Driver-from-Jython-td7142.html > ), > but it seems, that no one have really done this with Spark. > It looks like performance gain from using jython can be huge - you wouldn't > need to spawn PythonWorkers, all the code would be just executed inside > SparkExecutor JVM, using python code compiled to java bytecode. Do you > think > that's possible to achieve? Do you see any obvious obstacles? Of course, > jython doesn't have C extensions, but if one doesn't need them, then it > should fit here nicely. > > I'm willing to try to marry Spark with Jython and see how it goes. > > What do you think about this? > > > > > > -- > View this message in context: > http://apache-spark-developers-list.1001551.n3.nabble.com/How-to-speed-PySpark-to-match-Scala-Java-performance-tp10356.html > Sent from the Apache Spark Developers List mailing list archive at > Nabble.com. > > - > To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org > For additional commands, e-mail: dev-h...@spark.apache.org > >
Re: How to speed PySpark to match Scala/Java performance
Hi Reynold, In my project I want to use Python API too. When you mention DF's are we talking about pandas or this is something internal to spark py api. If you could elaborate a bit on this or point me to alternate documentation. Thanks much --sasha On Thu, Jan 29, 2015 at 4:12 PM, Reynold Xin wrote: > Once the data frame API is released for 1.3, you can write your thing in > Python and get the same performance. It can't express everything, but for > basic things like projection, filter, join, aggregate and simple numeric > computation, it should work pretty well. > > > On Thu, Jan 29, 2015 at 12:45 PM, rtshadow > > wrote: > > > Hi, > > > > In my company, we've been trying to use PySpark to run ETLs on our data. > > Alas, it turned out to be terribly slow compared to Java or Scala API > > (which > > we ended up using to meet performance criteria). > > > > To be more quantitative, let's consider simple case: > > I've generated test file (848MB): /seq 1 1 > /tmp/test/ > > > > and tried to run simple computation on it, which includes three steps: > read > > -> multiply each row by 2 -> take max > > Code in python: /sc.textFile("/tmp/test").map(lambda x: x * 2).max()/ > > Code in scala: /sc.textFile("/tmp/test").map(x => x * 2).max()/ > > > > Here are the results of this simple benchmark: > > CPython - 59s > > PyPy - 26s > > Scala version - 7s > > > > I didn't dig into what exactly contributes to execution times of CPython > / > > PyPy, but it seems that serialization / deserialization, when sending > data > > to the worker may be the issue. > > I know some guys already have been asking about using Jython > > ( > > > http://apache-spark-developers-list.1001551.n3.nabble.com/Jython-importing-pyspark-td8654.html#a8658 > > , > > > > > http://apache-spark-developers-list.1001551.n3.nabble.com/PySpark-Driver-from-Jython-td7142.html > > ), > > but it seems, that no one have really done this with Spark. > > It looks like performance gain from using jython can be huge - you > wouldn't > > need to spawn PythonWorkers, all the code would be just executed inside > > SparkExecutor JVM, using python code compiled to java bytecode. Do you > > think > > that's possible to achieve? Do you see any obvious obstacles? Of course, > > jython doesn't have C extensions, but if one doesn't need them, then it > > should fit here nicely. > > > > I'm willing to try to marry Spark with Jython and see how it goes. > > > > What do you think about this? > > > > > > > > > > > > -- > > View this message in context: > > > http://apache-spark-developers-list.1001551.n3.nabble.com/How-to-speed-PySpark-to-match-Scala-Java-performance-tp10356.html > > Sent from the Apache Spark Developers List mailing list archive at > > Nabble.com. > > > > - > > To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org > > For additional commands, e-mail: dev-h...@spark.apache.org > > > > > -- Aleksandar Kacanski
Re: How to speed PySpark to match Scala/Java performance
It is something like this: https://issues.apache.org/jira/browse/SPARK-5097 On the master branch, we have a Pandas like API already. On Thu, Jan 29, 2015 at 4:31 PM, Sasha Kacanski wrote: > Hi Reynold, > In my project I want to use Python API too. > When you mention DF's are we talking about pandas or this is something > internal to spark py api. > If you could elaborate a bit on this or point me to alternate > documentation. > Thanks much --sasha > > On Thu, Jan 29, 2015 at 4:12 PM, Reynold Xin wrote: > >> Once the data frame API is released for 1.3, you can write your thing in >> Python and get the same performance. It can't express everything, but for >> basic things like projection, filter, join, aggregate and simple numeric >> computation, it should work pretty well. >> >> >> On Thu, Jan 29, 2015 at 12:45 PM, rtshadow < >> pastuszka.przemys...@gmail.com> >> wrote: >> >> > Hi, >> > >> > In my company, we've been trying to use PySpark to run ETLs on our data. >> > Alas, it turned out to be terribly slow compared to Java or Scala API >> > (which >> > we ended up using to meet performance criteria). >> > >> > To be more quantitative, let's consider simple case: >> > I've generated test file (848MB): /seq 1 1 > /tmp/test/ >> > >> > and tried to run simple computation on it, which includes three steps: >> read >> > -> multiply each row by 2 -> take max >> > Code in python: /sc.textFile("/tmp/test").map(lambda x: x * 2).max()/ >> > Code in scala: /sc.textFile("/tmp/test").map(x => x * 2).max()/ >> > >> > Here are the results of this simple benchmark: >> > CPython - 59s >> > PyPy - 26s >> > Scala version - 7s >> > >> > I didn't dig into what exactly contributes to execution times of >> CPython / >> > PyPy, but it seems that serialization / deserialization, when sending >> data >> > to the worker may be the issue. >> > I know some guys already have been asking about using Jython >> > ( >> > >> http://apache-spark-developers-list.1001551.n3.nabble.com/Jython-importing-pyspark-td8654.html#a8658 >> > , >> > >> > >> http://apache-spark-developers-list.1001551.n3.nabble.com/PySpark-Driver-from-Jython-td7142.html >> > ), >> > but it seems, that no one have really done this with Spark. >> > It looks like performance gain from using jython can be huge - you >> wouldn't >> > need to spawn PythonWorkers, all the code would be just executed inside >> > SparkExecutor JVM, using python code compiled to java bytecode. Do you >> > think >> > that's possible to achieve? Do you see any obvious obstacles? Of course, >> > jython doesn't have C extensions, but if one doesn't need them, then it >> > should fit here nicely. >> > >> > I'm willing to try to marry Spark with Jython and see how it goes. >> > >> > What do you think about this? >> > >> > >> > >> > >> > >> > -- >> > View this message in context: >> > >> http://apache-spark-developers-list.1001551.n3.nabble.com/How-to-speed-PySpark-to-match-Scala-Java-performance-tp10356.html >> > Sent from the Apache Spark Developers List mailing list archive at >> > Nabble.com. >> > >> > - >> > To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org >> > For additional commands, e-mail: dev-h...@spark.apache.org >> > >> > >> > > > > -- > Aleksandar Kacanski >
Re: How to speed PySpark to match Scala/Java performance
thanks for quick reply, I will check the link. Hopefully, with conversion to py3, or 3.4 we could take advantage of asyncio and other cool new stuff ... On Thu, Jan 29, 2015 at 7:41 PM, Reynold Xin wrote: > It is something like this: > https://issues.apache.org/jira/browse/SPARK-5097 > > On the master branch, we have a Pandas like API already. > > > On Thu, Jan 29, 2015 at 4:31 PM, Sasha Kacanski > wrote: > >> Hi Reynold, >> In my project I want to use Python API too. >> When you mention DF's are we talking about pandas or this is something >> internal to spark py api. >> If you could elaborate a bit on this or point me to alternate >> documentation. >> Thanks much --sasha >> >> On Thu, Jan 29, 2015 at 4:12 PM, Reynold Xin wrote: >> >>> Once the data frame API is released for 1.3, you can write your thing in >>> Python and get the same performance. It can't express everything, but for >>> basic things like projection, filter, join, aggregate and simple numeric >>> computation, it should work pretty well. >>> >>> >>> On Thu, Jan 29, 2015 at 12:45 PM, rtshadow < >>> pastuszka.przemys...@gmail.com> >>> wrote: >>> >>> > Hi, >>> > >>> > In my company, we've been trying to use PySpark to run ETLs on our >>> data. >>> > Alas, it turned out to be terribly slow compared to Java or Scala API >>> > (which >>> > we ended up using to meet performance criteria). >>> > >>> > To be more quantitative, let's consider simple case: >>> > I've generated test file (848MB): /seq 1 1 > /tmp/test/ >>> > >>> > and tried to run simple computation on it, which includes three steps: >>> read >>> > -> multiply each row by 2 -> take max >>> > Code in python: /sc.textFile("/tmp/test").map(lambda x: x * 2).max()/ >>> > Code in scala: /sc.textFile("/tmp/test").map(x => x * 2).max()/ >>> > >>> > Here are the results of this simple benchmark: >>> > CPython - 59s >>> > PyPy - 26s >>> > Scala version - 7s >>> > >>> > I didn't dig into what exactly contributes to execution times of >>> CPython / >>> > PyPy, but it seems that serialization / deserialization, when sending >>> data >>> > to the worker may be the issue. >>> > I know some guys already have been asking about using Jython >>> > ( >>> > >>> http://apache-spark-developers-list.1001551.n3.nabble.com/Jython-importing-pyspark-td8654.html#a8658 >>> > , >>> > >>> > >>> http://apache-spark-developers-list.1001551.n3.nabble.com/PySpark-Driver-from-Jython-td7142.html >>> > ), >>> > but it seems, that no one have really done this with Spark. >>> > It looks like performance gain from using jython can be huge - you >>> wouldn't >>> > need to spawn PythonWorkers, all the code would be just executed inside >>> > SparkExecutor JVM, using python code compiled to java bytecode. Do you >>> > think >>> > that's possible to achieve? Do you see any obvious obstacles? Of >>> course, >>> > jython doesn't have C extensions, but if one doesn't need them, then it >>> > should fit here nicely. >>> > >>> > I'm willing to try to marry Spark with Jython and see how it goes. >>> > >>> > What do you think about this? >>> > >>> > >>> > >>> > >>> > >>> > -- >>> > View this message in context: >>> > >>> http://apache-spark-developers-list.1001551.n3.nabble.com/How-to-speed-PySpark-to-match-Scala-Java-performance-tp10356.html >>> > Sent from the Apache Spark Developers List mailing list archive at >>> > Nabble.com. >>> > >>> > - >>> > To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org >>> > For additional commands, e-mail: dev-h...@spark.apache.org >>> > >>> > >>> >> >> >> >> -- >> Aleksandar Kacanski >> > > -- Aleksandar Kacanski