GitHub user manishgupta88 opened a pull request:
https://github.com/apache/carbondata/pull/2210
[CARBONDATA-2381] Improve compaction performance by filling batch result in
columnar format and performing IO at blocklet level
Problem: Compaction performance is slow as compared to data load
Analysis:
1. During compaction result filling is done in row format. Due to this as
the number of columns increases the dimension and measure data filling time
increases. This happens because in row filling we are not able to take
advantage of OS cacheable buffers as we continuously read data for next column.
2. As compaction uses a page level reader flow wherein both IO and
uncompression is done at page level, the IO and uncompression time increases in
this model.
Solution:
1. Implement a columnar format filling data structure for compaction
process for filling dimension and measure data.
2. Perform IO at blocklet level and uncompression at page level.
Be sure to do all of the following checklist to help us incorporate
your contribution quickly and easily:
- [ ] Any interfaces changed?
- [ ] Any backward compatibility impacted?
- [ ] Document update required?
- [ ] Testing done
Please provide details on
- Whether new unit test cases have been added or why no new tests
are required?
- How it is tested? Please attach test report.
- Is it a performance related change? Please attach the performance
test report.
- Any additional information to help reviewers in testing this
change.
- [ ] For large changes, please consider breaking it into sub-tasks under
an umbrella JIRA.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/manishgupta88/carbondata
compaction_slow_fix_1.4
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/carbondata/pull/2210.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #2210
----
commit 857e5d2a75eaf16457023f5a04a2e420bb8cbff2
Author: manishgupta88 <tomanishgupta18@...>
Date: 2018-04-23T07:06:39Z
Problem: Compaction performance is slow as compared to data load
Analysis:
1. During compaction result filling is done in row format. Due to this as
the number of columns increases the dimension and measure data filling time
increases. This happens because in row filling we are not able to take
advantage of OS cacheable buffers as we continuously read data for next column.
2. As compaction uses a page level reader flow wherein both IO and
uncompression is done at page level, the IO and uncompression time
increases in this model.
Solution:
1. Implement a columnar format filling data structure for compaction
process for filling dimension and measure data.
2. Perform IO at blocklet level and uncompression at page level.
----
---
