[HIVE-4440] SMB Operator spills to disk like it's 1999

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.12.0
• Assignee: Gunther Hagleitner
• Reporter: Gunther Hagleitner
• Created: 29/Apr/13 02:49
• Updated: 19/May/14 08:50
• Resolved: 15/May/13 21:11

Description

I was recently looking into some performance issue with a query that used SMB join and was running really slow. Turns out that the SMB join by default caches only 100 values per key before spilling to disk. That seems overly conservative to me. Changing the parameter resulted in a ~5x speedup - quite significant.

The parameter is: hive.mapjoin.bucket.cache.size

Which right now is only used the SMB Operator as far as I can tell.

The parameter was introduced originally (3 yrs ago) for the map join operator (looks like pre-SMB) and set to 100 to avoid OOM. That seems to have been in a different context though where you had to avoid running out of memory with the cached hash table in the same process, I think.

Two things I'd like to propose:

a) Rename it to what it does: hive.smbjoin.cache.rows
b) Set it to something less restrictive: 10000

If you string together a 5 table smb join with a map join and a map-side group by aggregation you might still run out of memory, but the renamed parameter should be easier to find and reduce. For most queries, I would think that 10000 is still a reasonable number to cache (On the reduce side we use 25000 for shuffle joins).

Issue Links

• depends upon: HIVE-4439 Remove unused join configuration parameter: hive.mapjoin.cache.numrows