b5c60bcdca
## What changes were proposed in this pull request? This patch speeds up group-by aggregates by around 3-5x by leveraging an in-memory `AggregateHashMap` (please see https://github.com/apache/spark/pull/12161), an append-only aggregate hash map that can act as a 'cache' for extremely fast key-value lookups while evaluating aggregates (and fall back to the `BytesToBytesMap` if a given key isn't found). Architecturally, it is backed by a power-of-2-sized array for index lookups and a columnar batch that stores the key-value pairs. The index lookups in the array rely on linear probing (with a small number of maximum tries) and use an inexpensive hash function which makes it really efficient for a majority of lookups. However, using linear probing and an inexpensive hash function also makes it less robust as compared to the `BytesToBytesMap` (especially for a large number of keys or even for certain distribution of keys) and requires us to fall back on the latter for correctness. ## How was this patch tested? Java HotSpot(TM) 64-Bit Server VM 1.8.0_73-b02 on Mac OS X 10.11.4 Intel(R) Core(TM) i7-4960HQ CPU 2.60GHz Aggregate w keys: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative ------------------------------------------------------------------------------------------- codegen = F 2124 / 2204 9.9 101.3 1.0X codegen = T hashmap = F 1198 / 1364 17.5 57.1 1.8X codegen = T hashmap = T 369 / 600 56.8 17.6 5.8X Author: Sameer Agarwal <sameer@databricks.com> Closes #12345 from sameeragarwal/tungsten-aggregate-integration. |
||
---|---|---|
.. | ||
compatibility/src/test/scala/org/apache/spark/sql/hive/execution | ||
src | ||
pom.xml |