Details
-
New Feature
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
-
None
Description
The Hive query execution engine currently processes one row at a time. A single row of data goes through all the operators before the next row can be processed. This mode of processing is very inefficient in terms of CPU usage. Research has demonstrated that this yields very low instructions per cycle [MonetDB X100]. Also currently Hive heavily relies on lazy deserialization and data columns go through a layer of object inspectors that identify column type, deserialize data and determine appropriate expression routines in the inner loop. These layers of virtual method calls further slow down the processing.
This work will add support for vectorized query execution to Hive, where, instead of individual rows, batches of about a thousand rows at a time are processed. Each column in the batch is represented as a vector of a primitive data type. The inner loop of execution scans these vectors very fast, avoiding method calls, deserialization, unnecessary if-then-else, etc. This substantially reduces CPU time used, and gives excellent instructions per cycle (i.e. improved processor pipeline utilization). See the attached design specification for more details.
Attachments
Attachments
Issue Links
- incorporates
-
HIVE-5584 Write initial user documentation for vectorized query on Hive Wiki
-
- Resolved
-
- relates to
-
HIVE-10179 Optimization for SIMD instructions in Hive
-
- Open
-
-
-
- In Progress
-
