[HIVE-7664] VectorizedBatchUtil.addRowToBatchFrom is not optimized for Vectorized execution and takes 25% CPU

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 0.13.1
• Tags: hive
• Assignee: Matt McCline
• Reporter: Mostafa Mokhtar
• Created: 08/Aug/14 21:25
• Updated: 22/Jul/18 08:38
• Resolved: 22/Jul/18 08:38

Description

In a Group by heavy vectorized Reducer vertex 25% of CPU is spent in VectorizedBatchUtil.addRowToBatchFrom().

Looked at the code of VectorizedBatchUtil.addRowToBatchFrom and it looks like it wasn't optimized for Vectorized processing.

addRowToBatchFrom is called for every row and for each row and every column in the batch getPrimitiveCategory is called to figure the type of each column, column types are stored in a HashMap, for VectorGroupByOperator columns types won't change between batches, so column types shouldn't be looked up for every row.

I recommend storing the column type in StructObjectInspector so that other components can leverage this optimization.

Also addRowToBatchFrom has a case statement for every row and every column used for type casting I recommend encapsulating the type logic in templatized methods.

Stack Trace	Sample Count	Percentage(%)
VectorizedBatchUtil.addRowToBatchFrom	86	26.543
   AbstractPrimitiveObjectInspector.getPrimitiveCategory()	34	10.494
   LazyBinaryStructObjectInspector.getStructFieldData	25	7.716
   StandardStructObjectInspector.getStructFieldData	4	1.235

The query used :

select 
    ss_sold_date_sk
from
    store_sales
where
    ss_sold_date between '1998-01-01' and '1998-06-01'
group by ss_item_sk , ss_customer_sk , ss_sold_date_sk
having sum(ss_list_price) > 50000000000000;

Issue Links

• is superceded by: HIVE-9937 LLAP: Vectorized Field-By-Field Serialize / Deserialize to support new Vectorized Map Join
• links to: review board