I have been using PySpark with Ipython lately on my server with 24 CPUs and 32GB RAM. Its running only on one machine. In my process, I want to collect huge amount of data as is give in below code:
train_dataRDD = (train.map(lambda x:getTagsAndText(x))
.filter(lambda x:x[-1]!=[])
.flatMap(lambda (x,text,tags): [(tag,(x,text)) for tag in tags])
.groupByKey()
.mapValues(list))
When I do
training_data = train_dataRDD.collectAsMap()
It gives me outOfMemory Error. Java heap Space
. Also, I can not perform any operations on Spark after this error as it looses connection with Java. It gives Py4JNetworkError: Cannot connect to the java server
.
It looks like heap space is small. How can I set it to bigger limits?
EDIT:
Things that I tried before running:
sc._conf.set('spark.executor.memory','32g').set('spark.driver.memory','32g').set('spark.driver.maxResultsSize','0')
I changed the spark options as per the documentation here(if you do ctrl-f and search for spark.executor.extraJavaOptions) : http://spark.apache.org/docs/1.2.1/configuration.html
It says that I can avoid OOMs by setting spark.executor.memory option. I did the same thing but it seem not be working.
After trying out loads of configuration parameters, I found that there is only one need to be changed to enable more Heap space and i.e. spark.driver.memory
.
sudo vim $SPARK_HOME/conf/spark-defaults.conf
#uncomment the spark.driver.memory and change it according to your use. I changed it to below
spark.driver.memory 15g
# press : and then wq! to exit vim editor
Close your existing spark application and re run it. You will not encounter this error again. :)