Toby Potter created SPARK-7682:
----------------------------------
Summary: Size of distributed grids still limited by cPickle
Key: SPARK-7682
URL: https://issues.apache.org/jira/browse/SPARK-7682
Project: Spark
Issue Type: Bug
Components: PySpark
Affects Versions: 1.3.1
Environment: Redhat Enterprise Linux 6.5, Spark 1.3.1 standalone in
cluster mode, 2 nodes with 64 GB spark slaves, Python 2.7.6
Reporter: Toby Potter
Priority: Minor
I'm trying to explore the possibilities of writing a fault-tolerant distributed
computing engine for multidimensional arrays. I'm finding that the Python
cPickle serializer is limiting the size of Numpy arrays that I can distribute
over the cluster.
My example code is below
#!/usr/bin/env python
#Python app to use spark
from pyspark import SparkContext, SparkConf
import numpy
appName="Spark Test App"
# Create a spark context
conf = SparkConf().setAppName(appName)
# Set memory
conf = SparkConf().set("spark.executor.memory", "32g")
sc = SparkContext(conf=conf)
# Make array
grid=numpy.zeros((1024,1024,1024))
# Now parallelise and persist the data
rdd = sc.parallelize([("srcw", grid)])
# Make the data persist in memory
rdd_rdd.persist()
When I run the code I get the following error
Traceback (most recent call last):
File "test_app.py", line 20, in <module>
rdd = sc.parallelize([("srcw", grid)])
File "/spark/1.3.1/python/pyspark/context.py", line 341, in parallelize
serializer.dump_stream(c, tempFile)
File "/spark/1.3.1/python/pyspark/serializers.py", line 208, in dump_stream
self.serializer.dump_stream(self._batched(iterator), stream)
File "/spark/1.3.1/python/pyspark/serializers.py", line 127, in dump_stream
self._write_with_length(obj, stream)
File "/spark/1.3.1/python/pyspark/serializers.py", line 137, in
_write_with_length
serialized = self.dumps(obj)
File "/spark/1.3.1/python/pyspark/serializers.py", line 403, in dumps
return cPickle.dumps(obj, 2)
SystemError: error return without exception set
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
