feat: direct submission variable timeout

Add mechanism to increase direct submission timeout up to a maximum
value when no new submissions were made since last sleep.

This should help in workloads that have delays between iterations larger
than current direct submission controller timeout.

Related-To: NEO-7878

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
This commit is contained in:
Dominik Dabek
2023-04-06 12:14:02 +00:00
committed by Compute-Runtime-Automation
parent 51b8dc66a3
commit 411ed1c643
6 changed files with 150 additions and 29 deletions

View File

@@ -339,6 +339,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionDisableMonitorFence, -1, "Disabl
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionFlatRingBuffer, -1, "-1: default, 0: disable, 1: enable, Copies task command buffer directly into ring, implemented for immediate command lists only")
DECLARE_DEBUG_VARIABLE(int32_t, EnableDirectSubmissionController, -1, "Enable direct submission terminating after given timeout, -1: default, 0: disabled, 1: enabled")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerTimeout, -1, "Set direct submission controller timeout, -1: default 5000 us, >=0: timeout in us")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerMaxTimeout, -1, "Set direct submission controller max timeout - timeout will increase up to given value, -1: default 5000 us, >=0: max timeout in us")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerDivisor, -1, "Set direct submission controller timeout divider, -1: default 1, >0: divider value")
DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionForceLocalMemoryStorageMode, -1, "Force local memory storage for command/ring/semaphore buffer, -1: default - for all engines, 0: disabled, 1: for multiOsContextCapable engine, 2: for all engines")
DECLARE_DEBUG_VARIABLE(int32_t, EnableRingSwitchTagUpdateWa, -1, "-1: default, 0 - disable, 1 - enable. If enabled, completionFences wont be updated if ring is not running.")

View File

@@ -19,11 +19,14 @@ namespace NEO {
DirectSubmissionController::DirectSubmissionController() {
if (DebugManager.flags.DirectSubmissionControllerTimeout.get() != -1) {
timeout = DebugManager.flags.DirectSubmissionControllerTimeout.get();
timeout = std::chrono::microseconds{DebugManager.flags.DirectSubmissionControllerTimeout.get()};
}
if (DebugManager.flags.DirectSubmissionControllerDivisor.get() != -1) {
timeoutDivisor = DebugManager.flags.DirectSubmissionControllerDivisor.get();
}
if (DebugManager.flags.DirectSubmissionControllerMaxTimeout.get() != -1) {
maxTimeout = std::chrono::microseconds{DebugManager.flags.DirectSubmissionControllerMaxTimeout.get()};
}
directSubmissionControllingThread = Thread::create(controlDirectSubmissionsState, reinterpret_cast<void *>(this));
};
@@ -74,7 +77,7 @@ void *DirectSubmissionController::controlDirectSubmissionsState(void *self) {
void DirectSubmissionController::checkNewSubmissions() {
std::lock_guard<std::mutex> lock(this->directSubmissionsMutex);
bool shouldRecalculateTimeout = false;
for (auto &directSubmission : this->directSubmissions) {
auto csr = directSubmission.first;
auto &state = directSubmission.second;
@@ -87,18 +90,26 @@ void DirectSubmissionController::checkNewSubmissions() {
auto lock = csr->obtainUniqueOwnership();
csr->stopDirectSubmission();
state.isStopped = true;
shouldRecalculateTimeout = true;
}
} else {
state.isStopped = false;
state.taskCount = taskCount;
}
}
if (shouldRecalculateTimeout) {
this->recalculateTimeout();
}
}
void DirectSubmissionController::sleep() {
std::this_thread::sleep_for(std::chrono::microseconds(this->timeout));
}
SteadyClock::time_point DirectSubmissionController::getCpuTimestamp() {
return SteadyClock::now();
}
void DirectSubmissionController::adjustTimeout(CommandStreamReceiver *csr) {
if (EngineHelpers::isCcs(csr->getOsContext().getEngineType())) {
for (size_t subDeviceIndex = 0u; subDeviceIndex < csr->getOsContext().getDeviceBitfield().size(); ++subDeviceIndex) {
@@ -114,4 +125,16 @@ void DirectSubmissionController::adjustTimeout(CommandStreamReceiver *csr) {
}
}
void DirectSubmissionController::recalculateTimeout() {
const auto now = this->getCpuTimestamp();
const auto timeSinceLastTerminate = std::chrono::duration_cast<std::chrono::microseconds>(now - this->lastTerminateCpuTimestamp);
DEBUG_BREAK_IF(timeSinceLastTerminate.count() < 0);
if (timeSinceLastTerminate.count() > this->timeout.count() &&
timeSinceLastTerminate.count() <= this->maxTimeout.count()) {
const auto newTimeout = std::chrono::duration_cast<std::chrono::microseconds>(timeSinceLastTerminate * 1.5);
this->timeout = newTimeout.count() < this->maxTimeout.count() ? newTimeout : this->maxTimeout;
}
this->lastTerminateCpuTimestamp = now;
}
} // namespace NEO

View File

@@ -12,6 +12,7 @@
#include <array>
#include <atomic>
#include <chrono>
#include <memory>
#include <mutex>
#include <unordered_map>
@@ -21,8 +22,11 @@ class MemoryManager;
class CommandStreamReceiver;
class Thread;
using SteadyClock = std::chrono::steady_clock;
class DirectSubmissionController {
public:
static constexpr size_t defaultTimeout = 5'000;
DirectSubmissionController();
virtual ~DirectSubmissionController();
@@ -42,8 +46,10 @@ class DirectSubmissionController {
static void *controlDirectSubmissionsState(void *self);
void checkNewSubmissions();
MOCKABLE_VIRTUAL void sleep();
MOCKABLE_VIRTUAL SteadyClock::time_point getCpuTimestamp();
void adjustTimeout(CommandStreamReceiver *csr);
void recalculateTimeout();
uint32_t maxCcsCount = 1u;
std::array<uint32_t, DeviceBitfield().size()> ccsCount = {};
@@ -54,7 +60,9 @@ class DirectSubmissionController {
std::atomic_bool keepControlling = true;
std::atomic_bool runControlling = false;
int timeout = 5000;
SteadyClock::time_point lastTerminateCpuTimestamp{};
std::chrono::microseconds maxTimeout{defaultTimeout};
std::chrono::microseconds timeout{defaultTimeout};
int timeoutDivisor = 1;
};
} // namespace NEO

View File

@@ -512,3 +512,4 @@ ForceDummyBlitWa = -1
DetectIndirectAccessInKernel = -1
OptimizeIoqBarriersHandling = -1
AllocateSharedAllocationsInHeapExtended = 0
DirectSubmissionControllerMaxTimeout = -1

View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2019-2022 Intel Corporation
* Copyright (C) 2019-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -16,6 +16,8 @@ struct DirectSubmissionControllerMock : public DirectSubmissionController {
using DirectSubmissionController::directSubmissions;
using DirectSubmissionController::directSubmissionsMutex;
using DirectSubmissionController::keepControlling;
using DirectSubmissionController::lastTerminateCpuTimestamp;
using DirectSubmissionController::maxTimeout;
using DirectSubmissionController::timeout;
using DirectSubmissionController::timeoutDivisor;
@@ -24,6 +26,11 @@ struct DirectSubmissionControllerMock : public DirectSubmissionController {
this->sleepCalled = true;
}
SteadyClock::time_point getCpuTimestamp() override {
return cpuTimestamp;
}
SteadyClock::time_point cpuTimestamp{};
bool sleepCalled = false;
};
} // namespace NEO

View File

@@ -22,7 +22,7 @@ TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllerTimeoutWhen
DirectSubmissionControllerMock controller;
EXPECT_EQ(controller.timeout, 14);
EXPECT_EQ(controller.timeout.count(), 14);
}
TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllertimeoutDivisorWhenCreateObjectThentimeoutDivisorIsEqualWithDebugFlag) {
@@ -124,6 +124,87 @@ TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllerWithStarted
controller.directSubmissionControllingThread.reset();
}
TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllerAndDivisorDisabledWhenIncreaseTimeoutEnabledThenTimeoutIsIncreased) {
DebugManagerStateRestore restorer;
DebugManager.flags.DirectSubmissionControllerMaxTimeout.set(200'000);
DebugManager.flags.DirectSubmissionControllerDivisor.set(1);
MockExecutionEnvironment executionEnvironment;
executionEnvironment.prepareRootDeviceEnvironments(1);
executionEnvironment.initializeMemoryManager();
DeviceBitfield deviceBitfield(1);
MockCommandStreamReceiver csr(executionEnvironment, 0, deviceBitfield);
std::unique_ptr<OsContext> osContext(OsContext::create(nullptr, 0, 0,
EngineDescriptorHelper::getDefaultDescriptor({aub_stream::ENGINE_CCS, EngineUsage::Regular},
PreemptionMode::ThreadGroup, deviceBitfield)));
csr.setupContext(*osContext.get());
DirectSubmissionControllerMock controller;
controller.keepControlling.store(false);
controller.directSubmissionControllingThread->join();
controller.directSubmissionControllingThread.reset();
controller.registerDirectSubmission(&csr);
{
csr.taskCount.store(1u);
controller.checkNewSubmissions();
EXPECT_FALSE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 1u);
auto previousTimestamp = controller.lastTerminateCpuTimestamp;
controller.cpuTimestamp += std::chrono::microseconds(5'000);
controller.checkNewSubmissions();
EXPECT_EQ(std::chrono::duration_cast<std::chrono::microseconds>(controller.lastTerminateCpuTimestamp - previousTimestamp).count(), 5'000);
EXPECT_TRUE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 1u);
EXPECT_EQ(controller.timeout.count(), 5'000);
EXPECT_EQ(controller.maxTimeout.count(), 200'000);
}
{
csr.taskCount.store(2u);
controller.checkNewSubmissions();
EXPECT_FALSE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 2u);
auto previousTimestamp = controller.lastTerminateCpuTimestamp;
controller.cpuTimestamp += std::chrono::microseconds(5'500);
controller.checkNewSubmissions();
EXPECT_EQ(std::chrono::duration_cast<std::chrono::microseconds>(controller.lastTerminateCpuTimestamp - previousTimestamp).count(), 5'500);
EXPECT_TRUE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 2u);
EXPECT_EQ(controller.timeout.count(), 8'250);
}
{
csr.taskCount.store(3u);
controller.checkNewSubmissions();
EXPECT_FALSE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 3u);
auto previousTimestamp = controller.lastTerminateCpuTimestamp;
controller.cpuTimestamp += controller.maxTimeout;
controller.checkNewSubmissions();
EXPECT_EQ(std::chrono::duration_cast<std::chrono::microseconds>(controller.lastTerminateCpuTimestamp - previousTimestamp).count(), controller.maxTimeout.count());
EXPECT_TRUE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 3u);
EXPECT_EQ(controller.timeout.count(), controller.maxTimeout.count());
}
{
controller.timeout = std::chrono::microseconds(5'000);
csr.taskCount.store(4u);
controller.checkNewSubmissions();
EXPECT_FALSE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 4u);
auto previousTimestamp = controller.lastTerminateCpuTimestamp;
controller.cpuTimestamp += controller.maxTimeout * 2;
controller.checkNewSubmissions();
EXPECT_EQ(std::chrono::duration_cast<std::chrono::microseconds>(controller.lastTerminateCpuTimestamp - previousTimestamp).count(), controller.maxTimeout.count() * 2);
EXPECT_TRUE(controller.directSubmissions[&csr].isStopped);
EXPECT_EQ(controller.directSubmissions[&csr].taskCount, 4u);
EXPECT_EQ(controller.timeout.count(), 5'000);
}
controller.unregisterDirectSubmission(&csr);
}
TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllerWithNotStartedControllingWhenShuttingDownThenNoHang) {
DirectSubmissionControllerMock controller;
EXPECT_NE(controller.directSubmissionControllingThread.get(), nullptr);
@@ -176,22 +257,22 @@ TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllerWhenRegiste
controller.directSubmissionControllingThread->join();
controller.directSubmissionControllingThread.reset();
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr3);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr1);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr2);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr4);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.unregisterDirectSubmission(&csr);
controller.unregisterDirectSubmission(&csr1);
@@ -280,37 +361,37 @@ TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllerWhenRegiste
controller.directSubmissionControllingThread->join();
controller.directSubmissionControllingThread.reset();
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr5);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr1);
EXPECT_EQ(controller.timeout, 1250);
EXPECT_EQ(controller.timeout.count(), 1'250);
controller.registerDirectSubmission(&csr2);
EXPECT_EQ(controller.timeout, 312);
EXPECT_EQ(controller.timeout.count(), 312);
controller.registerDirectSubmission(&csr4);
EXPECT_EQ(controller.timeout, 312);
EXPECT_EQ(controller.timeout.count(), 312);
controller.registerDirectSubmission(&csr6);
EXPECT_EQ(controller.timeout, 312);
EXPECT_EQ(controller.timeout.count(), 312);
controller.registerDirectSubmission(&csr7);
EXPECT_EQ(controller.timeout, 312);
EXPECT_EQ(controller.timeout.count(), 312);
controller.registerDirectSubmission(&csr9);
EXPECT_EQ(controller.timeout, 312);
EXPECT_EQ(controller.timeout.count(), 312);
controller.registerDirectSubmission(&csr8);
EXPECT_EQ(controller.timeout, 78);
EXPECT_EQ(controller.timeout.count(), 78);
controller.registerDirectSubmission(&csr10);
EXPECT_EQ(controller.timeout, 78);
EXPECT_EQ(controller.timeout.count(), 78);
controller.unregisterDirectSubmission(&csr);
controller.unregisterDirectSubmission(&csr1);
@@ -363,22 +444,22 @@ TEST(DirectSubmissionControllerTests, givenDirectSubmissionControllerDirectSubmi
controller.directSubmissionControllingThread->join();
controller.directSubmissionControllingThread.reset();
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr3);
EXPECT_EQ(controller.timeout, 5000);
EXPECT_EQ(controller.timeout.count(), 5'000);
controller.registerDirectSubmission(&csr1);
EXPECT_EQ(controller.timeout, 1000);
EXPECT_EQ(controller.timeout.count(), 1'000);
controller.registerDirectSubmission(&csr2);
EXPECT_EQ(controller.timeout, 200);
EXPECT_EQ(controller.timeout.count(), 200);
controller.registerDirectSubmission(&csr4);
EXPECT_EQ(controller.timeout, 200);
EXPECT_EQ(controller.timeout.count(), 200);
controller.unregisterDirectSubmission(&csr);
controller.unregisterDirectSubmission(&csr1);