birkeh/o3de
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Makes the recorded benchmark more stable

Browse Source 
Signed-off-by: Esteban Papp <81431996+amznestebanpapp@users.noreply.github.com>
This commit is contained in:
Esteban Papp
2021-12-08 13:44:15 -08:00
parent bb33f906d2
commit 0d66278ef7
3 changed files with 135 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Code/Framework/AzCore/AzCore/Memory/AllocatorBase.cpp
+16 -7
View File
@@ -11,7 +11,7 @@
using namespace AZ;
#define RECORDING_ENABLED 1
#define RECORDING_ENABLED 0
#if RECORDING_ENABLED
@@ -44,18 +44,22 @@ namespace
}
};
struct AllocatorOperation
#pragma pack(push, 1)
struct alignas(1) AllocatorOperation
{
enum OperationType : unsigned int
enum OperationType : size_t
{
ALLOCATE,
DEALLOCATE
};
OperationType m_type: 1;
unsigned int m_size : 28; // Can represent up to 256Mb requests
unsigned int m_alignment : 7; // Can represent up to 128 alignment
unsigned int m_recordId : 28; // Can represent up to 256M simultaneous requests, we reuse ids
size_t m_size : 28; // Can represent up to 256Mb requests
size_t m_alignment : 7; // Can represent up to 128 alignment
size_t m_recordId : 28; // Can represent up to 256M simultaneous requests, we reuse ids
};
#pragma pack(pop)
static_assert(sizeof(AllocatorOperation) == 8);
static AZStd::mutex s_operationsMutex = {};
static constexpr size_t s_maxNumberOfAllocationsToRecord = 16384;
@@ -76,7 +80,12 @@ namespace
if (s_operationCounter == s_allocationOperationCount)
{
AZ::IO::SystemFile file;
file.Open("memoryrecordings.bin", AZ::IO::SystemFile::OpenMode::SF_OPEN_APPEND);
int mode = AZ::IO::SystemFile::OpenMode::SF_OPEN_APPEND | AZ::IO::SystemFile::OpenMode::SF_OPEN_WRITE_ONLY;
if (!file.Exists("memoryrecordings.bin"))
{
mode |= AZ::IO::SystemFile::OpenMode::SF_OPEN_CREATE;
}
file.Open("memoryrecordings.bin", mode);
if (file.IsOpen())
{
file.Write(&s_operations, sizeof(AllocatorOperation) * s_allocationOperationCount);
Code/Framework/AzCore/Tests/Memory/AllocatorBenchmarkRecordings.bin
+2 -2
View File
@@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d0f148441ce120b303896618cec364b5afb6f8b911b4785ec6358cfe8467cf7a
size 368640
oid sha256:281ba03e79ecba90b313a0b17bdba87c57d76b504b6e38d579b5eabd995902cc
size 245760
Code/Framework/AzCore/Tests/Memory/AllocatorBenchmarks.cpp
+117 -69
View File
@@ -390,115 +390,159 @@ namespace Benchmark
}
}
};
template<typename TAllocator>
class RecordedAllocationBenchmarkFixture : public AllocatorBenchmarkFixture<TAllocator>
{
using base = AllocatorBenchmarkFixture<TAllocator>;
using TestAllocatorType = typename base::TestAllocatorType;
struct AllocatorOperation
template<typename TAllocator>
class RecordedAllocationBenchmarkFixture : public ::benchmark::Fixture
{
using TestAllocatorType = TestAllocatorWrapper<TAllocator>;
virtual void internalSetUp()
{
enum OperationType : unsigned int
TestAllocatorType::SetUp();
}
void internalTearDown()
{
TestAllocatorType::TearDown();
}
#pragma pack(push, 1)
struct alignas(1) AllocatorOperation
{
enum OperationType : size_t
{
ALLOCATE,
DEALLOCATE
};
OperationType m_type : 1;
unsigned int m_size : 28; // Can represent up to 256Mb requests
unsigned int m_alignment : 7; // Can represent up to 128 alignment
unsigned int m_recordId : 28; // Can represent up to 256M simultaneous requests, we reuse ids
size_t m_size : 28; // Can represent up to 256Mb requests
size_t m_alignment : 7; // Can represent up to 128 alignment
size_t m_recordId : 28; // Can represent up to 256M simultaneous requests, we reuse ids
};
#pragma pack(pop)
static_assert(sizeof(AllocatorOperation) == 8);
public:
void SetUp(const ::benchmark::State&) override
{
internalSetUp();
}
void SetUp(::benchmark::State&) override
{
internalSetUp();
}
void TearDown(const ::benchmark::State&) override
{
internalTearDown();
}
void TearDown(::benchmark::State&) override
{
internalTearDown();
}
void Benchmark(benchmark::State& state)
{
for (auto _ : state)
{
state.PauseTiming();
AZStd::unordered_map<unsigned int, void*> pointerRemapping;
AZStd::unordered_map<size_t, void*> pointerRemapping;
constexpr size_t allocationOperationCount = 5 * 1024;
AZStd::array<AllocatorOperation, allocationOperationCount> m_operations = {};
[[maybe_unused]] const size_t operationSize = sizeof(AllocatorOperation);
AZ::IO::SystemFile file;
AZ::IO::FixedMaxPathString filePath = AZ::Utils::GetExecutableDirectory();
filePath += "/Tests/AzCore/Memory/AllocatorBenchmarkRecordings.bin";
if (!file.Open(filePath.c_str(), AZ::IO::SystemFile::OpenMode::SF_OPEN_READ_ONLY))
{
return;
}
size_t elementsRead = file.Read(sizeof(AllocatorOperation) * allocationOperationCount, &m_operations);
size_t totalElementsRead = elementsRead;
const size_t processMemoryBaseline = Platform::GetProcessMemoryUsageBytes();
size_t totalAllocationSize = 0;
size_t itemsProcessed = 0;
while (elementsRead > 0)
for (size_t i = 0; i < 100; ++i) // replay the recording, this way we can keep a smaller recording
{
for (size_t operationIndex = 0; operationIndex < elementsRead; ++operationIndex)
AZ::IO::SystemFile file;
AZ::IO::FixedMaxPathString filePath = AZ::Utils::GetExecutableDirectory();
filePath += "/Tests/AzCore/Memory/AllocatorBenchmarkRecordings.bin";
if (!file.Open(filePath.c_str(), AZ::IO::SystemFile::OpenMode::SF_OPEN_READ_ONLY))
{
const AllocatorOperation& operation = m_operations[operationIndex];
if (operation.m_type == AllocatorOperation::ALLOCATE)
return;
}
size_t elementsRead =
file.Read(sizeof(AllocatorOperation) * allocationOperationCount, &m_operations) / sizeof(AllocatorOperation);
itemsProcessed += elementsRead;
while (elementsRead > 0)
{
for (size_t operationIndex = 0; operationIndex < elementsRead; ++operationIndex)
{
const auto it = pointerRemapping.emplace(operation.m_recordId, nullptr);
if (it.second) // otherwise already allocated
const AllocatorOperation& operation = m_operations[operationIndex];
if (operation.m_type == AllocatorOperation::ALLOCATE)
{
state.ResumeTiming();
void* ptr = TestAllocatorType::Allocate(operation.m_size, operation.m_alignment);
state.PauseTiming();
totalAllocationSize += operation.m_size;
it.first->second = ptr;
}
else
{
// Doing a resize, dont account for this memory change, this operation is rare and we dont have
// the size of the previous allocation
state.ResumeTiming();
TestAllocatorType::Resize(it.first->second, operation.m_size);
state.PauseTiming();
}
}
else // AllocatorOperation::DEALLOCATE:
{
if (operation.m_recordId)
{
const auto ptrIt = pointerRemapping.find(operation.m_recordId);
if (ptrIt != pointerRemapping.end())
const auto it = pointerRemapping.emplace(operation.m_recordId, nullptr);
if (it.second) // otherwise already allocated
{
totalAllocationSize -= operation.m_size;
state.ResumeTiming();
TestAllocatorType::DeAllocate(ptrIt->second, /*operation.m_size*/ 0); // size is not correct after a resize, a 0 size deals with it
void* ptr = TestAllocatorType::Allocate(operation.m_size, operation.m_alignment);
state.PauseTiming();
totalAllocationSize += operation.m_size;
it.first->second = ptr;
}
else
{
// Doing a resize, dont account for this memory change, this operation is rare and we dont have
// the size of the previous allocation
state.ResumeTiming();
TestAllocatorType::Resize(it.first->second, operation.m_size);
state.PauseTiming();
pointerRemapping.erase(ptrIt);
}
}
else // deallocate(nullptr) are recorded
else // AllocatorOperation::DEALLOCATE:
{
// Just to account of the call of deallocate(nullptr);
state.ResumeTiming();
TestAllocatorType::DeAllocate(nullptr, /*operation.m_size*/ 0);
state.PauseTiming();
if (operation.m_recordId)
{
const auto ptrIt = pointerRemapping.find(operation.m_recordId);
if (ptrIt != pointerRemapping.end())
{
totalAllocationSize -= operation.m_size;
state.ResumeTiming();
TestAllocatorType::DeAllocate(
ptrIt->second,
/*operation.m_size*/ 0); // size is not correct after a resize, a 0 size deals with it
state.PauseTiming();
pointerRemapping.erase(ptrIt);
}
}
else // deallocate(nullptr) are recorded
{
// Just to account of the call of deallocate(nullptr);
state.ResumeTiming();
TestAllocatorType::DeAllocate(nullptr, /*operation.m_size*/ 0);
state.PauseTiming();
}
}
}
}
elementsRead = file.Read(sizeof(AllocatorOperation) * allocationOperationCount, &m_operations);
totalElementsRead += elementsRead;
elementsRead =
file.Read(sizeof(AllocatorOperation) * allocationOperationCount, &m_operations) / sizeof(AllocatorOperation);
itemsProcessed += elementsRead;
}
file.Close();
// Deallocate the remainder (since we stopped the recording middle-game)(there are leaks as well)
for (const auto& pointerMapping : pointerRemapping)
{
state.ResumeTiming();
TestAllocatorType::DeAllocate(pointerMapping.second);
state.PauseTiming();
}
itemsProcessed += pointerRemapping.size();
pointerRemapping.clear();
}
file.Close();
state.counters[s_counterAllocatorMemory] = benchmark::Counter(static_cast<double>(TestAllocatorType::NumAllocatedBytes()), benchmark::Counter::kDefaults);
state.counters[s_counterProcessMemory] = benchmark::Counter(static_cast<double>(Platform::GetProcessMemoryUsageBytes() - processMemoryBaseline), benchmark::Counter::kDefaults);
state.counters[s_counterBenchmarkMemory] = benchmark::Counter(static_cast<double>(totalAllocationSize), benchmark::Counter::kDefaults);
state.SetItemsProcessed(totalElementsRead);
state.SetItemsProcessed(itemsProcessed);
// Deallocate the remainder (since we stopped the recording middle-game)(there are leaks as well)
for (const auto& pointerMapping : pointerRemapping)
{
TestAllocatorType::DeAllocate(pointerMapping.second);
}
pointerRemapping.clear();
TestAllocatorType::GarbageCollect();
}
}
@@ -514,8 +558,7 @@ namespace Benchmark
}
static void RecordedRunRanges(benchmark::internal::Benchmark* b)
{
b->Arg(1);
b->Iterations(100);
b->Iterations(1);
}
// For threaded ranges, run just 200, multi-threaded will already multiply by thread
@@ -547,6 +590,11 @@ namespace Benchmark
BM_REGISTER_TEMPLATE(RecordedAllocationBenchmarkFixture, TESTNAME, ALLOCATORTYPE)->Apply(RecordedRunRanges); \
}
/// Warm up benchmark used to prepare the OS for allocations. Most OS keep allocations for a process somehow
/// reserved. So the first allocations run always get a bigger impact in a process. This warm up allocator runs
/// all the benchmarks and is just used for the the next allocators to report more consistent results.
BM_REGISTER_ALLOCATOR(WarmUpAllocator, TestRawMallocAllocator);
BM_REGISTER_ALLOCATOR(RawMallocAllocator, TestRawMallocAllocator);
BM_REGISTER_ALLOCATOR(MallocSchemaAllocator, TestMallocSchemaAllocator);
BM_REGISTER_ALLOCATOR(HphaSchemaAllocator, TestHphaSchemaAllocator);