perf(swordfish): Parallel expression evaluation #3593
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
CodSpeed Performance ReportMerging #3593 will degrade performances by 15.97%Comparing Summary
Benchmarks breakdown
|
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #3593 +/- ##
==========================================
+ Coverage 77.80% 77.83% +0.02%
==========================================
Files 718 718
Lines 88176 88295 +119
==========================================
+ Hits 68607 68726 +119
Misses 19569 19569
|
colin-ho
commented
Dec 19, 2024
| @@ -46,7 +46,7 @@ impl Session { | |||
| let cfg = Arc::new(DaftExecutionConfig::default()); | |||
| let native_executor = NativeExecutor::from_logical_plan_builder(&optimized_plan)?; | |||
|
|
|||
| let mut result_stream = native_executor.run(&pset, cfg, None)?.into_stream(); | |||
| let mut result_stream = pin!(native_executor.run(&pset, cfg, None)?.into_stream()); | |||
There was a problem hiding this comment.
The kanal receiver doesn't provide a way to get an owned stream, so I manually made one using unfold, with a recv.await inside. Because of this await point I need to pin the stream.
colin-ho
commented
Dec 19, 2024
| @@ -36,7 +32,7 @@ pub(crate) fn create_ordering_aware_receiver_channel<T: Clone>( | |||
| ) -> (Vec<Sender<T>>, OrderingAwareReceiver<T>) { | |||
| match ordered { | |||
| true => { | |||
| let (senders, receiver) = (0..buffer_size).map(|_| create_channel::<T>(1)).unzip(); | |||
| let (senders, receiver) = (0..buffer_size).map(|_| create_channel::<T>(0)).unzip(); | |||
There was a problem hiding this comment.
Setting channel sizes to 0 can elide heap allocation.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Addresses: #3389. More generally, this PR optimizes for projections with many expressions, particularly memory intensive expressions like UDFs.
Problem:
Currently, swordfish parallelizes projections across morsels, with 1 CPU per morsel. However, if each projection has many memory intensive expressions, we could experience a massive inflation in memory because we will have many materialized morsels living in memory at once.
Proposed solution:
Instead, we can parallelize the expressions within the projection (but only for expressions that require compute). This way, we still have good CPU utilization, but we keep a lower number of materialized morsels in memory.
In the linked issue above, we see that a 128cpu machine will parallelize morsels across the cores, each doing multiple udfs, resulting in "317GB allocations and duration 351 secs".
This PR reduces that to 7.8GB peak memory and runtime of 66 seconds.
Notes:
asyncsend to asyncreceive was not respecting capacity constraints, and was allowing sends even though the receive did not happen. Moved over to https://github.com/fereidani/kanal, which worked much better.Todos for next time: