Skip to content

fix(core): drain GPU workers before unregister #409

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.

Sign up for GitHub

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

Closed
xiaguan wants to merge 1 commit into from
Closed

fix(core): drain GPU workers before unregister #409

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
220 changes: 167 additions & 53 deletions pegaflow-core/src/gpu_worker.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@ use std::sync::{Arc, mpsc as std_mpsc};
use cudarc::driver::{CudaContext, CudaStream};
use log::{debug, error, info, warn};
use logforth::diagnostic::ThreadLocalDiagnostic;
use parking_lot::Mutex;
use tokio::sync::{mpsc, oneshot};

use crate::EngineError;
Expand Down Expand Up @@ -112,11 +113,38 @@ pub(crate) struct SaveTask {
pub trace_ctx: Option<::fastrace::prelude::SpanContext>,
}

enum LoadCommand {
Run(LoadTask),
Drain(oneshot::Sender<()>),
}

enum SaveCommand {
Run(SaveTask),
Drain(oneshot::Sender<()>),
}

pub(crate) struct WorkerDrain {
load: oneshot::Receiver<()>,
save: oneshot::Receiver<()>,
}

impl WorkerDrain {
pub(crate) async fn wait(self) {
self.load
.await
.unwrap_or_else(|_| panic!("load worker exited before acknowledging drain barrier"));
self.save
.await
.unwrap_or_else(|_| panic!("save worker exited before acknowledging drain barrier"));
}
}

/// Per-GPU worker pool with dedicated load and save threads
pub(crate) struct GpuWorkerPool {
device_id: i32,
load_tx: mpsc::UnboundedSender<LoadTask>,
save_tx: mpsc::UnboundedSender<SaveTask>,
load_tx: mpsc::UnboundedSender<LoadCommand>,
save_tx: mpsc::UnboundedSender<SaveCommand>,
closed: Mutex<bool>,
}

impl GpuWorkerPool {
Expand Down Expand Up @@ -197,12 +225,20 @@ impl GpuWorkerPool {
device_id,
load_tx,
save_tx,
closed: Mutex::new(false),
})
}

/// Submit a load task (CPU -> GPU) - fire and forget
pub(crate) fn submit_load(&self, task: LoadTask) -> Result<(), EngineError> {
self.load_tx.send(task).map_err(|_| {
let closed = self.closed.lock();
if *closed {
return Err(EngineError::Storage(format!(
"GPU worker pool is closing for device {}",
self.device_id
)));
}
self.load_tx.send(LoadCommand::Run(task)).map_err(|_| {
EngineError::Storage(format!(
"Load worker channel closed for device {}",
self.device_id
Expand All @@ -225,12 +261,21 @@ impl GpuWorkerPool {
trace_ctx: ::fastrace::prelude::SpanContext::current_local_parent(),
};

self.save_tx.send(task).map_err(|_| {
EngineError::Storage(format!(
"Save worker channel closed for device {}",
self.device_id
))
})?;
{
let closed = self.closed.lock();
if *closed {
return Err(EngineError::Storage(format!(
"GPU worker pool is closing for device {}",
self.device_id
)));
}
self.save_tx.send(SaveCommand::Run(task)).map_err(|_| {
EngineError::Storage(format!(
"Save worker channel closed for device {}",
self.device_id
))
})?;
}

// Await the result (this is async, won't block tokio runtime)
reply_rx.await.map_err(|_| {
Expand All @@ -240,6 +285,31 @@ impl GpuWorkerPool {
))
})?
}

/// Reject new transfers and place one barrier behind every transfer that
/// was accepted before this call.
pub(crate) fn close_and_drain(&self) -> WorkerDrain {
let mut closed = self.closed.lock();
assert!(!*closed, "GPU worker pool closed twice");
*closed = true;

let (load_tx, load_rx) = oneshot::channel();
let (save_tx, save_rx) = oneshot::channel();
let load_send = self.load_tx.send(LoadCommand::Drain(load_tx));
let save_send = self.save_tx.send(SaveCommand::Drain(save_tx));
assert!(
load_send.is_ok(),
"load worker exited before accepting drain barrier"
);
assert!(
save_send.is_ok(),
"save worker exited before accepting drain barrier"
);
WorkerDrain {
load: load_rx,
save: save_rx,
}
}
}

struct WorkerRuntime {
Expand Down Expand Up @@ -300,18 +370,28 @@ fn init_worker(device_id: i32, transfer_mode: TransferMode) -> Result<WorkerRunt
/// Load worker thread main loop
fn load_worker_loop(
device_id: i32,
mut rx: mpsc::UnboundedReceiver<LoadTask>,
mut rx: mpsc::UnboundedReceiver<LoadCommand>,
runtime: WorkerRuntime,
) {
while let Some(task) = rx.blocking_recv() {
let LoadTask { layers, completion } = task;
let result = process_load_task(&layers, &runtime.stream, runtime.backend.as_ref());

if let Err(ref e) = result {
error!("Load task failed: device={device_id} error={e:?}");
core_metrics().load_failures.add(1, &[]);
while let Some(command) = rx.blocking_recv() {
match command {
LoadCommand::Run(task) => {
let LoadTask { layers, completion } = task;
let result = process_load_task(&layers, &runtime.stream, runtime.backend.as_ref());

if let Err(ref e) = result {
error!("Load task failed: device={device_id} error={e:?}");
core_metrics().load_failures.add(1, &[]);
}
completion.signal(result);
}
LoadCommand::Drain(done) => {
runtime.stream.synchronize().unwrap_or_else(|err| {
panic!("load stream drain failed for device {device_id}: {err:?}")
});
let _ = done.send(());
}
}
completion.signal(result);
}

info!("Load worker shutting down: device={}", device_id);
Expand All @@ -320,25 +400,35 @@ fn load_worker_loop(
/// Save worker thread main loop
fn save_worker_loop(
device_id: i32,
mut rx: mpsc::UnboundedReceiver<SaveTask>,
mut rx: mpsc::UnboundedReceiver<SaveCommand>,
runtime: WorkerRuntime,
) {
while let Some(task) = rx.blocking_recv() {
let SaveTask {
layers,
reply,
#[cfg(feature = "tracing")]
trace_ctx,
} = task;
let result = process_save_task(
&layers,
&runtime.stream,
runtime.backend.as_ref(),
#[cfg(feature = "tracing")]
trace_ctx,
)
.map(|()| layers);
let _ = reply.send(result);
while let Some(command) = rx.blocking_recv() {
match command {
SaveCommand::Run(task) => {
let SaveTask {
layers,
reply,
#[cfg(feature = "tracing")]
trace_ctx,
} = task;
let result = process_save_task(
&layers,
&runtime.stream,
runtime.backend.as_ref(),
#[cfg(feature = "tracing")]
trace_ctx,
)
.map(|()| layers);
let _ = reply.send(result);
}
SaveCommand::Drain(done) => {
runtime.stream.synchronize().unwrap_or_else(|err| {
panic!("save stream drain failed for device {device_id}: {err:?}")
});
let _ = done.send(());
}
}
}

info!("Save worker shutting down: device={}", device_id);
Expand Down Expand Up @@ -414,6 +504,46 @@ fn build_copy_descs(layers: &[LayerTransferData]) -> Result<(Vec<CopyDesc>, usiz
Ok((copies, total_bytes))
}

fn wait_for_transfer(
stream: &Arc<CudaStream>,
submitted: Result<(), String>,
) -> Result<(), EngineError> {
if let Err(submit_error) = submitted {
stream.synchronize().map_err(|sync_error| {
EngineError::Storage(format!(
"Failed to synchronize after transfer submission error ({submit_error}): {sync_error:?}"
))
})?;
return Err(EngineError::Storage(submit_error));
}

match stream.record_event(None) {
Ok(event) => match event.synchronize() {
Ok(()) => Ok(()),
Err(event_error) => {
stream.synchronize().map_err(|stream_error| {
EngineError::Storage(format!(
"Event synchronization failed ({event_error:?}); stream synchronization also failed: {stream_error:?}"
))
})?;
Err(EngineError::Storage(format!(
"Failed to synchronize event: {event_error:?}"
)))
}
},
Err(record_error) => {
stream.synchronize().map_err(|sync_error| {
EngineError::Storage(format!(
"Failed to synchronize after event record failure ({record_error:?}): {sync_error:?}"
))
})?;
Err(EngineError::Storage(format!(
"Failed to record event: {record_error:?}"
)))
}
}
}

/// Process a load task: copy blocks from CPU pinned memory to GPU for multiple
/// layers. All layers and segments are collected into one descriptor batch,
/// handed to a single backend (memcpy or kernel), then synchronized once.
Expand All @@ -430,15 +560,7 @@ fn process_load_task(

let (copies, total_bytes) = build_copy_descs(layers)?;

backend.h2d(&copies, stream).map_err(EngineError::Storage)?;

// Wait for all transfers to complete
let event = stream
.record_event(None)
.map_err(|e| EngineError::Storage(format!("Failed to record event: {e:?}")))?;
event
.synchronize()
.map_err(|e| EngineError::Storage(format!("Failed to synchronize: {e:?}")))?;
wait_for_transfer(stream, backend.h2d(&copies, stream))?;

let elapsed = start.elapsed();
let bandwidth_gbps = if elapsed.as_secs_f64() > 0.0 {
Expand Down Expand Up @@ -483,15 +605,7 @@ fn process_save_task(

let (copies, total_bytes) = build_copy_descs(layers)?;

backend.d2h(&copies, stream).map_err(EngineError::Storage)?;

// Single synchronization for all layers
let event = stream
.record_event(None)
.map_err(|e| EngineError::Storage(format!("Failed to record event: {e:?}")))?;
event
.synchronize()
.map_err(|e| EngineError::Storage(format!("Failed to synchronize: {e:?}")))?;
wait_for_transfer(stream, backend.d2h(&copies, stream))?;

let elapsed = start.elapsed();
let bandwidth_gbps = if elapsed.as_secs_f64() > 0.0 {
Expand Down
Loading
Loading