From ab9ff98c693490db3dd5bc2bb89c774699d86ad3 Mon Sep 17 00:00:00 2001
From: Stefan Boberg <stefan@boberg.org>
Date: Thu, 2 Apr 2026 18:54:47 +0200
Subject: Improve compute scheduler throughput and reduce lock contention

- Parallelize per-runner SubmitActions via WorkerThreadPool when
  multiple runners are active
- Make remote submissions non-blocking: dispatch to pool and return
  immediately, marking actions as Submitting so the scheduler won't
  re-pick them; failures transition to Failed for auto-retry
- Downgrade m_ActionMapLock from exclusive to shared in
  SchedulePendingActions and move sort/resize outside the lock
- Release m_RunnersLock before submitting by snapshotting runners
  as Ref copies, avoiding holding the lock during HTTP I/O
- Add per-runner submit timing warning (>500ms) with host context
- Add GetDisplayName() virtual to FunctionRunner for diagnostics
---
 src/zencompute/runners/functionrunner.cpp | 100 +++++++++++++++++++++++++-----
 1 file changed, 83 insertions(+), 17 deletions(-)

(limited to 'src/zencompute/runners/functionrunner.cpp')

diff --git a/src/zencompute/runners/functionrunner.cpp b/src/zencompute/runners/functionrunner.cpp
index 67e12b84e..f8a54b3f8 100644
--- a/src/zencompute/runners/functionrunner.cpp
+++ b/src/zencompute/runners/functionrunner.cpp
@@ -6,9 +6,13 @@
 
 #	include <zencore/compactbinary.h>
 #	include <zencore/filesystem.h>
+#	include <zencore/logging.h>
+#	include <zencore/timer.h>
 #	include <zencore/trace.h>
+#	include <zencore/workthreadpool.h>
 
 #	include <fmt/format.h>
+#	include <future>
 #	include <vector>
 
 namespace zen::compute {
@@ -118,23 +122,34 @@ std::vector<SubmitResult>
 BaseRunnerGroup::SubmitActions(const std::vector<Ref<RunnerAction>>& Actions)
 {
 	ZEN_TRACE_CPU("BaseRunnerGroup::SubmitActions");
-	RwLock::SharedLockScope _(m_RunnersLock);
 
-	const int RunnerCount = gsl::narrow<int>(m_Runners.size());
+	// Snapshot runners and query capacity under the lock, then release
+	// before submitting — HTTP submissions to remote runners can take
+	// hundreds of milliseconds and we must not hold m_RunnersLock during I/O.
 
-	if (RunnerCount == 0)
-	{
-		return std::vector<SubmitResult>(Actions.size(), SubmitResult{.IsAccepted = false, .Reason = "No runners available"});
-	}
+	std::vector<Ref<FunctionRunner>>			Runners;
+	std::vector<size_t>							Capacities;
+	std::vector<std::vector<Ref<RunnerAction>>> PerRunnerActions;
+	size_t										TotalCapacity = 0;
 
-	// Query capacity per runner and compute total
-	std::vector<size_t> Capacities(RunnerCount);
-	size_t				TotalCapacity = 0;
+	m_RunnersLock.WithSharedLock([&] {
+		const int RunnerCount = gsl::narrow<int>(m_Runners.size());
+		Runners.assign(m_Runners.begin(), m_Runners.end());
+		Capacities.resize(RunnerCount);
+		PerRunnerActions.resize(RunnerCount);
 
-	for (int i = 0; i < RunnerCount; ++i)
+		for (int i = 0; i < RunnerCount; ++i)
+		{
+			Capacities[i] = Runners[i]->QueryCapacity();
+			TotalCapacity += Capacities[i];
+		}
+	});
+
+	const int RunnerCount = gsl::narrow<int>(Runners.size());
+
+	if (RunnerCount == 0)
 	{
-		Capacities[i] = m_Runners[i]->QueryCapacity();
-		TotalCapacity += Capacities[i];
+		return std::vector<SubmitResult>(Actions.size(), SubmitResult{.IsAccepted = false, .Reason = "No runners available"});
 	}
 
 	if (TotalCapacity == 0)
@@ -143,9 +158,8 @@ BaseRunnerGroup::SubmitActions(const std::vector<Ref<RunnerAction>>& Actions)
 	}
 
 	// Distribute actions across runners proportionally to their available capacity
-	std::vector<std::vector<Ref<RunnerAction>>> PerRunnerActions(RunnerCount);
-	std::vector<size_t>							ActionRunnerIndex(Actions.size());
-	size_t										ActionIdx = 0;
+	std::vector<size_t> ActionRunnerIndex(Actions.size());
+	size_t				ActionIdx = 0;
 
 	for (int i = 0; i < RunnerCount; ++i)
 	{
@@ -176,14 +190,66 @@ BaseRunnerGroup::SubmitActions(const std::vector<Ref<RunnerAction>>& Actions)
 		}
 	}
 
-	// Submit batches per runner
+	// Submit batches per runner — in parallel when a worker pool is available
+
 	std::vector<std::vector<SubmitResult>> PerRunnerResults(RunnerCount);
 
+	int ActiveRunnerCount = 0;
 	for (int i = 0; i < RunnerCount; ++i)
 	{
 		if (!PerRunnerActions[i].empty())
 		{
-			PerRunnerResults[i] = m_Runners[i]->SubmitActions(PerRunnerActions[i]);
+			++ActiveRunnerCount;
+		}
+	}
+
+	static constexpr uint64_t SubmitWarnThresholdMs = 500;
+
+	auto SubmitToRunner = [&](int RunnerIndex) {
+		Stopwatch Timer;
+
+		PerRunnerResults[RunnerIndex] = Runners[RunnerIndex]->SubmitActions(PerRunnerActions[RunnerIndex]);
+
+		uint64_t ElapsedMs = Timer.GetElapsedTimeMs();
+		if (ElapsedMs >= SubmitWarnThresholdMs)
+		{
+			ZEN_WARN("submit of {} actions to '{}' took {}ms",
+					 PerRunnerActions[RunnerIndex].size(),
+					 Runners[RunnerIndex]->GetDisplayName(),
+					 ElapsedMs);
+		}
+	};
+
+	if (m_WorkerPool && ActiveRunnerCount > 1)
+	{
+		std::vector<std::future<void>> Futures(RunnerCount);
+
+		for (int i = 0; i < RunnerCount; ++i)
+		{
+			if (!PerRunnerActions[i].empty())
+			{
+				std::packaged_task<void()> Task([&SubmitToRunner, i]() { SubmitToRunner(i); });
+
+				Futures[i] = m_WorkerPool->EnqueueTask(std::move(Task), WorkerThreadPool::EMode::EnableBacklog);
+			}
+		}
+
+		for (int i = 0; i < RunnerCount; ++i)
+		{
+			if (Futures[i].valid())
+			{
+				Futures[i].get();
+			}
+		}
+	}
+	else
+	{
+		for (int i = 0; i < RunnerCount; ++i)
+		{
+			if (!PerRunnerActions[i].empty())
+			{
+				SubmitToRunner(i);
+			}
 		}
 	}
 
-- 
cgit v1.2.3