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

fix: recurse less in streaming shared sinks #13930

Merged
merged 1 commit into from
Jan 23, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 9 additions & 4 deletions crates/polars-pipe/src/pipeline/convert.rs
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@ use crate::executors::sinks::*;
use crate::executors::{operators, sources};
use crate::expressions::PhysicalPipedExpr;
use crate::operators::{Operator, Sink as SinkTrait, Source};
use crate::pipeline::dispatcher::SinkNode;
use crate::pipeline::PipeLine;

fn exprs_to_physical<F>(
Expand Down Expand Up @@ -668,7 +669,7 @@ where
let operator_offset = operator_objects.len();
operator_objects.extend(operators);

let sink_nodes = sink_nodes
let sinks = sink_nodes
.into_iter()
.map(|(offset, node, shared_count)| {
// ensure that shared sinks are really shared
Expand All @@ -685,16 +686,20 @@ where
Entry::Occupied(entry) => entry.get().split(0),
}
};

Ok((offset + operator_offset, node, sink, shared_count))
Ok(SinkNode::new(
sink,
shared_count,
offset + operator_offset,
node,
))
})
.collect::<PolarsResult<Vec<_>>>()?;

Ok(PipeLine::new(
source_objects,
operator_objects,
operator_nodes,
sink_nodes,
sinks,
operator_offset,
verbose,
))
Expand Down
103 changes: 65 additions & 38 deletions crates/polars-pipe/src/pipeline/dispatcher.rs
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,44 @@ use crate::operators::{
};
use crate::pipeline::morsels_per_sink;

pub(super) struct SinkNode {
pub sinks: Vec<Box<dyn Sink>>,
/// when that hits 0, the sink will finalize
pub shared_count: Rc<RefCell<u32>>,
initial_shared_count: u32,
/// - offset in the operators vec
/// at that point the sink should be called.
/// the pipeline will first call the operators on that point and then
/// push the result in the sink.
pub operator_end: usize,
pub node: Node,
}

impl SinkNode {
pub fn new(
sink: Box<dyn Sink>,
shared_count: Rc<RefCell<u32>>,
operator_end: usize,
node: Node,
) -> Self {
let n_threads = morsels_per_sink();
let sinks = (0..n_threads).map(|i| sink.split(i)).collect();
let initial_shared_count = *shared_count.borrow();
SinkNode {
sinks,
initial_shared_count,
shared_count,
operator_end,
node,
}
}

// Only the first node of a shared sink should recurse. The others should return.
fn allow_recursion(&self) -> bool {
self.initial_shared_count == *self.shared_count.borrow()
}
}

/// A pipeline consists of:
///
/// - 1. One or more sources.
Expand Down Expand Up @@ -69,8 +107,7 @@ pub struct PipeLine {
/// - shared_count
/// when that hits 0, the sink will finalize
/// - node of the sink
#[allow(clippy::type_complexity)]
sinks: Vec<(usize, Rc<RefCell<u32>>, Vec<Box<dyn Sink>>)>,
sinks: Vec<SinkNode>,
/// are used to identify the sink shared with other pipeline branches
sink_nodes: Vec<Node>,
/// Other branch of the pipeline/tree that must be executed
Expand All @@ -86,12 +123,11 @@ pub struct PipeLine {

impl PipeLine {
#[allow(clippy::type_complexity)]
pub fn new(
pub(super) fn new(
sources: Vec<Box<dyn Source>>,
operators: Vec<Box<dyn Operator>>,
operator_nodes: Vec<Node>,
// (offset, node (for identification), sink, shared_counter)
sink_and_nodes: Vec<(usize, Node, Box<dyn Sink>, Rc<RefCell<u32>>)>,
sinks: Vec<SinkNode>,
operator_offset: usize,
verbose: bool,
) -> PipeLine {
Expand All @@ -100,19 +136,8 @@ impl PipeLine {
// we only do that in the sinks itself.
let n_threads = morsels_per_sink();

let sink_nodes = sinks.iter().map(|s| s.node).collect();
// We split so that every thread gets an operator
let sink_nodes = sink_and_nodes.iter().map(|(_, node, _, _)| *node).collect();
let sinks = sink_and_nodes
.into_iter()
.map(|(offset, _, sink, shared_count)| {
(
offset,
shared_count,
(0..n_threads).map(|i| sink.split(i)).collect(),
)
})
.collect();

// every index maps to a chain of operators than can be pushed as a pipeline for one thread
let operators = (0..n_threads)
.map(|i| operators.iter().map(|op| op.split(i)).collect())
Expand Down Expand Up @@ -142,11 +167,11 @@ impl PipeLine {
sources,
operators,
vec![],
vec![(
operators_len,
Node::default(),
vec![SinkNode::new(
sink,
Rc::new(RefCell::new(1)),
operators_len,
Node::default(),
)],
0,
verbose,
Expand Down Expand Up @@ -345,9 +370,7 @@ impl PipeLine {
// we don't want to run the rest of the pipelines and we finalize early
let mut sink_finished = false;

for (i, (operator_end, shared_count, mut sink)) in
std::mem::take(&mut self.sinks).into_iter().enumerate()
{
for (i, mut sink) in std::mem::take(&mut self.sinks).into_iter().enumerate() {
for src in &mut std::mem::take(&mut self.sources) {
let mut next_batches = src.get_batches(ec)?;

Expand All @@ -357,10 +380,10 @@ impl PipeLine {

let (sink_result, next_batches2) = self.par_process_chunks(
chunks,
&mut sink,
&mut sink.sinks,
ec,
operator_start,
operator_end,
sink.operator_end,
src,
)?;
next_batches = next_batches2;
Expand All @@ -374,31 +397,35 @@ impl PipeLine {

// Before we reduce we also check if we should continue.
ec.execution_state.should_stop()?;
let allow_recursion = sink.allow_recursion();

// The sinks have taken all chunks thread locally, now we reduce them into a single
// result sink.
let mut reduced_sink = POOL
.install(|| {
sink.into_par_iter().reduce_with(|mut a, mut b| {
sink.sinks.into_par_iter().reduce_with(|mut a, mut b| {
a.combine(&mut *b);
a
})
})
.unwrap();
operator_start = operator_end;
operator_start = sink.operator_end;

let mut shared_sink_count = {
let mut shared_sink_count = shared_count.borrow_mut();
let mut shared_sink_count = sink.shared_count.borrow_mut();
*shared_sink_count -= 1;
*shared_sink_count
};

while shared_sink_count > 0 && !sink_finished {
let mut pipeline = pipeline_q.borrow_mut().pop_front().unwrap();
let (count, mut sink) =
pipeline.run_pipeline_no_finalize(ec, pipeline_q.clone())?;
reduced_sink.combine(sink.as_mut());
shared_sink_count = count;
// Prevent very deep recursion. Only the outer callee can pop and run.
if allow_recursion {
while shared_sink_count > 0 && !sink_finished {
let mut pipeline = pipeline_q.borrow_mut().pop_front().unwrap();
let (count, mut sink) =
pipeline.run_pipeline_no_finalize(ec, pipeline_q.clone())?;
reduced_sink.combine(sink.as_mut());
shared_sink_count = count;
}
}

if i != last_i {
Expand Down Expand Up @@ -491,18 +518,18 @@ impl Debug for PipeLine {
let mut fmt = String::new();
let mut start = 0usize;
fmt.push_str(self.sources[0].fmt());
for (offset_end, _, sink) in &self.sinks {
for sink in &self.sinks {
fmt.push_str(" -> ");
// take operators of a single thread
let ops = &self.operators[0];
// slice the pipeline
let ops = &ops[start..*offset_end];
let ops = &ops[start..sink.operator_end];
for op in ops {
fmt.push_str(op.fmt());
fmt.push_str(" -> ")
}
start = *offset_end;
fmt.push_str(sink[0].fmt())
start = sink.operator_end;
fmt.push_str(sink.sinks[0].fmt())
}
write!(f, "{fmt}")
}
Expand Down