xuchuanyin created CARBONDATA-1805:
--------------------------------------
Summary: Optimize pruning for dictionary loading
Key: CARBONDATA-1805
URL:
https://issues.apache.org/jira/browse/CARBONDATA-1805 Project: CarbonData
Issue Type: Improvement
Components: data-load, spark-integration
Reporter: xuchuanyin
Assignee: xuchuanyin
Fix For: 1.3.0
# SCENARIO
Recently I tried dictionary feature in Carbondata and found its dictionary generating phase in data loading is quite slow. My scenario is as below:
+ Input Data: 35.8GB CSV file with 199 columns and 126 Million lines
+ Dictionary columns: 3 columns each containing 19213,4,9 distinct values
The whole data loading consumes about 2.9min for dictionary generating and 4.6min for fact data loading -- about 39% of the time are spent on dictionary.
Having observed the nmon result, Ifound the CPU usage were quite high during the dictionary generating phase and the Disk, Network were quite normal.
# ANALYZE
After I went through the dictionary generating related code, I found Carbondata aleady prune non-dictionary columns before generating dictionary. But the problem is that `the pruning comes after data file reading`, this will cause some overhead, we can optimize it by `prune while reading data file`.
# RESOLVE
Refactor the `loadDataFrame` method in `GlobalDictionaryUtil`, only pruning the non-dictionary columns while reading the data file.
After implementing the above optimization, the dictionary generating costs only `29s` -- `about 6 times better than before`(2.9min), and the fact data loading costs the same as before(4.6min), about 10% of the time are spent on dictionary.
# NOTE
+ Currently only `load data file` will benefit from this optimization, while `load data frame` will not.
+ Before implementing this solution, I tried another solution -- cache dataframe of the data file, the performance was even worse -- the dictionary generating time was 5.6min.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
Locked
