Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Use tournament loser tree for k-way sort-merging, increase merge speed by 50% #4301

Merged
merged 1 commit into from
Nov 28, 2022

Conversation

richox
Copy link
Contributor

@richox richox commented Nov 21, 2022

Which issue does this PR close?

Closes #4300.

Rationale for this change

What changes are included in this PR?

Are these changes tested?

running benchmarks with cargo bench --bench merge:

merge i64               time:   [8.0959 ms 8.2001 ms 8.3260 ms]
                        change: [-58.629% -57.906% -57.138%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 11 outliers among 100 measurements (11.00%)
  4 (4.00%) high mild
  7 (7.00%) high severe

merge f64               time:   [8.2829 ms 8.3378 ms 8.4028 ms]
                        change: [-56.127% -55.475% -54.829%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
  5 (5.00%) high mild
  5 (5.00%) high severe

merge utf8 low cardinality
                        time:   [7.5418 ms 7.6425 ms 7.7538 ms]
                        change: [-65.897% -65.335% -64.703%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 13 outliers among 100 measurements (13.00%)
  7 (7.00%) high mild
  6 (6.00%) high severe

merge utf8 high cardinality
                        time:   [9.9438 ms 10.094 ms 10.257 ms]
                        change: [-52.043% -51.155% -50.249%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 8 outliers among 100 measurements (8.00%)
  4 (4.00%) high mild
  4 (4.00%) high severe

merge utf8 tuple        time:   [20.197 ms 20.493 ms 20.826 ms]
                        change: [-40.663% -39.505% -38.279%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
  6 (6.00%) high mild
  4 (4.00%) high severe

merge utf8 dictionary   time:   [5.1537 ms 5.2135 ms 5.2788 ms]
                        change: [-72.062% -71.616% -71.137%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 8 outliers among 100 measurements (8.00%)
  7 (7.00%) high mild
  1 (1.00%) high severe

merge utf8 dictionary tuple
                        time:   [7.9585 ms 8.0592 ms 8.1797 ms]
                        change: [-63.981% -63.324% -62.678%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 11 outliers among 100 measurements (11.00%)
  6 (6.00%) high mild
  5 (5.00%) high severe

merge mixed utf8 dictionary tuple
                        time:   [14.528 ms 14.669 ms 14.829 ms]
                        change: [-43.886% -43.201% -42.489%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
  5 (5.00%) high mild
  5 (5.00%) high severe

merge mixed tuple       time:   [17.824 ms 18.068 ms 18.333 ms]
                        change: [-39.957% -38.888% -37.748%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 5 outliers among 100 measurements (5.00%)
  4 (4.00%) high mild
  1 (1.00%) high severe

Are there any user-facing changes?

@github-actions github-actions bot added the core Core DataFusion crate label Nov 21, 2022
@alamb
Copy link
Contributor

alamb commented Nov 21, 2022

cc @tustvold

Copy link
Contributor

@tustvold tustvold left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks really good, thank you, I left some comments on how to make the implementation a little easier to follow, but I would be happy for this to go in as is.


if self.in_progress.len() == self.batch_size {
return Poll::Ready(Some(self.build_record_batch()));
let mut cmp_node = (num_streams + winner) / 2;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
let mut cmp_node = (num_streams + winner) / 2;
// Replace overall winner by walking tree of losers
let mut cmp_node = (num_streams + winner) / 2;

/// the loser nodes
loser_tree: Vec<usize>,

/// Identify whether the loser tree is adjusted
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
/// Identify whether the loser tree is adjusted
/// Identify whether the most recently yielded overall winner has been replaced
/// within the loser tree, a value of `false` indicates that they overall winner
/// has been yielded but the loser tree has not been updated

Or something to make it clearer what adjusted actually means.

FWIW a boolean of should_replace_winner or something might be clearer

self.current()
.cmp(&other.current())
.then_with(|| self.stream_idx.cmp(&other.stream_idx))
match (self.is_finished(), other.is_finished()) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
match (self.is_finished(), other.is_finished()) {
// Order finished cursors last
match (self.is_finished(), other.is_finished()) {


// Init all cursors and the loser tree in the first poll
if self.loser_tree.is_empty() {
// Ensure all non-exhausted streams have a cursor from which
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It might be easier to follow if this method were split into a method called init_loser_tree with a doc comment explaining what it does

self.cursor_finished[stream_idx] = true;
// Adjust the loser tree if necessary
if !self.loser_tree_adjusted {
let mut winner = self.loser_tree[0];
Copy link
Contributor

@tustvold tustvold Nov 21, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It might be easier to follow if this was moved into a method called replace_loser_tree_winner, perhaps with a link to this GIF - https://en.wikipedia.org/wiki/K-way_merge_algorithm#/media/File:Loser_tree_replacement_selection.gif

}
}
}
let min_cursor_idx = self.loser_tree[0];
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this could be made easier to follow if it were written along the lines of

let min_cursor = self.cursors[min_cursor_idx];
if min_cursor.is_finished() {
    // All streams are exhausted
    return Poll::Ready((!self.in_progress.is_empty()).then(|| self.build_record_batch()))
}

self.loser_tree_adjusted = false;
self.in_progress.push(...)
if self.in_progress.len() == self.batch_size {
    return Poll::Ready(Some(self.build_record_batch()));
}

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I couldn't make this work in #4407

Comment on lines +590 to +595
if challenger_win {
self.loser_tree[cmp_node] = winner;
winner = challenger;
} else {
self.loser_tree[cmp_node] = challenger;
}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
if challenger_win {
self.loser_tree[cmp_node] = winner;
winner = challenger;
} else {
self.loser_tree[cmp_node] = challenger;
}
if challenger_win {
self.loser_tree[cmp_node] = winner;
winner = challenger;
}

@alamb
Copy link
Contributor

alamb commented Nov 22, 2022

Thanks @richox @tustvold and @viirya -- this looks awesome. Let's wait to see if @richox would like to respond to comments.

@alamb
Copy link
Contributor

alamb commented Nov 26, 2022

If we don't hear back from @richox by Monday I plan to merge this PR and then create a follow on PR with the proposed changes from @tustvold and @viirya

@alamb alamb changed the title Use tournament loser tree for k-way sort-merging Use tournament loser tree for k-way sort-merging, increase merge speed by 50% Nov 28, 2022
@alamb
Copy link
Contributor

alamb commented Nov 28, 2022

I will have a follow on PR up shortly

@alamb alamb merged commit 0d334cf into apache:master Nov 28, 2022
@alamb
Copy link
Contributor

alamb commented Nov 28, 2022

Thanks again @richox @tustvold and @viirya

@ursabot
Copy link

ursabot commented Nov 28, 2022

Benchmark runs are scheduled for baseline = 52e198e and contender = 0d334cf. 0d334cf is a master commit associated with this PR. Results will be available as each benchmark for each run completes.
Conbench compare runs links:
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on ec2-t3-xlarge-us-east-2] ec2-t3-xlarge-us-east-2
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on test-mac-arm] test-mac-arm
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on ursa-i9-9960x] ursa-i9-9960x
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on ursa-thinkcentre-m75q] ursa-thinkcentre-m75q
Buildkite builds:
Supported benchmarks:
ec2-t3-xlarge-us-east-2: Supported benchmark langs: Python, R. Runs only benchmarks with cloud = True
test-mac-arm: Supported benchmark langs: C++, Python, R
ursa-i9-9960x: Supported benchmark langs: Python, R, JavaScript
ursa-thinkcentre-m75q: Supported benchmark langs: C++, Java

@alamb
Copy link
Contributor

alamb commented Nov 28, 2022

Follow on PR: #4407

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
core Core DataFusion crate
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Introduce tournament tree to achieve better k-way sort-merging
5 participants