nickva opened a new pull request, #6116:
URL: https://github.com/apache/couchdb/pull/6116
Switch from `1279` to `4096`
The chunk size is the limit at which B-tree nodes are split. `1279` default
was set 18+ [0] years ago by Damian after benchmarking on his personal laptop
with a spinning hard drive [1]. Nowadays most disks are SSDs and usually have a
4KB block size. We may read and write `1279` bytes but a modern disk will still
do the IO work to read 4KB blocks then throw the rest away. So it makes sense
to adjust it a bit to avoid wasting all that IO bandwidth for nothing.
Besides just the increased IO having a such a low default makes our B-tree
deeper which in turn generates even more IO and inflates our "active" size.
Raising the threshold to `4096` reduces the tree depth. For example on 200K doc
couch file Id tree went from depth 6 to 4 and the Seq tree from 4 to 3. This
also reduced the active size by about 25%. (To get an idea of what the B-tree
depths are there is a new depth function in `couch_btree` module and a bunch of
helper functions in `couch_debug` calling it).
The value of `4096` was first published in a research paper [2] by Chung et
al. on optimizing CouchDB for their embedded storage application. They saw a
20% TPS improvement with `4096` chunk size. To validate the setting we ran a k6
benchmark test on a large cluster with concurrent doc updates and reads.
Our test setup:
* 12 node cluster
* Q=32, N=3
* 10M docs, size 256B
* 1000 reads/s
* 200 writes/s
* 1000 writes/s via _bulk_docs
* 4 1 hour trials with each setting
Setting changed:
`btree_chunk_size` 1279 -> 4096
Results:
| Scenario | P50 (msec) | P90 (msec) |
|------------|---------------------------|------------------------|
| doc_get | 6.9 -> 5.1 (-27%) | 11.1 -> 6.4 (-43%) |
| doc_insert | 12.2 -> 8.3 (-32%) | 37.5 -> 10.9 (-71%) |
| doc_update | 14.1 -> 9.0 (-36%) | 44.4 -> 12.2 (-73%) |
| bulk_docs | 101 -> 64 (-37%) | 136 -> 83 (-39%) |
We do even better than what Chung et al. found. It could be because we now
have a B-tree block cache which favors shallow B-trees as it prioritizes nodes
closer to the root.
[0]
https://github.com/apache/couchdb/commit/10881200f1d6761d504badb2e12a16a38033c721
> Damien F. Katz committed on Jul 28, 2008
[1]
https://web.archive.org/web/20120315120536/http://ayende.com/blog/3609/more-couchdb-reading-btree-query-modify
> The ideal size of CHUNK_THRESHOLD was determined to be 0x4FF when timing
some
test runs on my MacBook Pro.
[2] https://arxiv.org/pdf/1912.11456 Performance Tuning and Scaling
Enterprise Blockchain Applications. Grant Chung et al.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
