Details
-
Bug
-
Status: Open
-
Minor
-
Resolution: Unresolved
-
0.20.1, 0.20.2, 0.21.0
-
None
-
None
-
None
-
Cloudera CDH3b3 (0.20.2+)
Description
I have MapReduce jobs that use a large amount of per-task memory, because the algorithm I'm using converges faster if more data is together on a node. I have my JVM heap size set at 3200 MB, and if I use the popular rule of thumb that io.sort.mb should be ~70% of that, I get 2240 MB. I rounded this down to 2048 MB, but map tasks crash with :
java.io.IOException: Invalid "io.sort.mb": 2048 at org.apache.hadoop.mapred.MapTask$MapOutputBuffer.<init>(MapTask.java:790) ...
MapTask.MapOutputBuffer implements its buffer with a byte[] of size io.sort.mb (in bytes), and is sanity checking the size before allocating the array. The problem is that Java arrays can't have more than 2^31 - 1 elements (even with a 64-bit JVM), and this is a limitation of the Java language specificiation itself. As memory and data sizes grow, this would seem to be a crippling limtiation of Java.
It would be nice if this ceiling were documented, and an error issued sooner, e.g. in jobtracker startup upon reading the config. Going forward, we may need to implement some array of arrays hack for large buffers.
Attachments
Issue Links
- is duplicated by
-
MAPREDUCE-5705 mapreduce.task.io.sort.mb hardcoded cap at 2047
- Reopened