Skip to content

Optimize rebatch kernel in Adsim #198

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
Kaiweitu wants to merge 1 commit into from
Closed

Optimize rebatch kernel in Adsim #198

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
106 changes: 85 additions & 21 deletions packages/adsim/src/cpp2/server/dwarfs/Rebatch.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -244,14 +244,16 @@ Rebatch::Rebatch(
size_t memory_pool_size_gb,
int prefetch_dist,
std::string input_str,
std::string distribution_file)
std::string distribution_file,
size_t num_pregenerated_batches)
: tensors_per_batch_(tensors_per_batch),
num_threads_(num_threads),
output_tensor_size_(output_tensor_size),
memory_pool_size_gb_(memory_pool_size_gb),
prefetch_dist_(prefetch_dist),
input_str_(std::move(input_str)),
distribution_file_(std::move(distribution_file)) {
distribution_file_(std::move(distribution_file)),
num_pregenerated_batches_(num_pregenerated_batches) {
LOG(INFO) << "Rebatch initialized with params: " << "tensors_per_batch="
<< tensors_per_batch_ << ", num_threads=" << num_threads_
<< ", output_tensor_size=" << output_tensor_size_
Expand Down Expand Up @@ -282,25 +284,69 @@ std::string Rebatch::init(
h_objs->set_shared_ptr(input_str_ + "_size_dist", size_distribution_);
h_objs->set_shared_ptr(input_str_ + "_memory_pool", input_memory_pool_);

// Pregenerate batches for performance optimization
pregenerateBatches();

std::string info = folly::to<std::string>(
"Rebatch: ", input_str_, " ", std::to_string(output_tensor_size_), "B");
info += " (distribution: " + distribution_file_ + ")";
info += " (memory_pool: " + std::to_string(memory_pool_size_gb_) + "GB)";
info +=
" (pregenerated_batches: " + std::to_string(num_pregenerated_batches_) +
")";
return info;
}

/**
* Pregenerate N batches to avoid on-the-fly generation during benchmark
* execution
*
* Creates pregenerated_batches_ vector with N TensorBatch instances that can be
* reused in a round-robin fashion during rebatch operations.
*/
void Rebatch::pregenerateBatches() {
LOG(INFO) << "Pregenerating " << num_pregenerated_batches_
<< " batches for performance optimization";

pregenerated_batches_.reserve(num_pregenerated_batches_);

for (size_t i = 0; i < num_pregenerated_batches_; ++i) {
auto batch = std::make_shared<TensorBatch>(
tensors_per_batch_,
*size_distribution_,
output_tensor_size_,
*input_memory_pool_);
pregenerated_batches_.push_back(std::move(batch));
}

LOG(INFO) << "Successfully pregenerated " << pregenerated_batches_.size()
<< " batches";
}

/**
* Get next pregenerated batch in round-robin fashion
*
* Returns a pregenerated batch from the pool, cycling through them in a
* thread-safe manner using atomic operations.
*/
std::shared_ptr<TensorBatch> Rebatch::getNextBatch() {
if (pregenerated_batches_.empty()) {
throw std::runtime_error("No pregenerated batches available");
}

size_t index = batch_index_.fetch_add(1) % pregenerated_batches_.size();
return pregenerated_batches_[index];
}

/**
* Single-threaded rebatch operation
*
* Creates a batch of tensors and performs sequential memory copy operations.
* Uses a pregenerated batch of tensors and performs sequential memory copy
* operations.
*/
int Rebatch::rebatch_single_thread() {
// Create a fresh batch of tensors for this operation
auto batch = std::make_unique<TensorBatch>(
tensors_per_batch_,
*size_distribution_,
output_tensor_size_,
*input_memory_pool_);
// Get a pregenerated batch for this operation (avoids on-the-fly generation)
auto batch = getNextBatch();

const auto& tensors = batch->getTensors();
const size_t numTensors = tensors.size();
Expand Down Expand Up @@ -341,10 +387,17 @@ int Rebatch::rebatch_single_thread() {

// Perform the memcpy operation
memcpy(output.get() + totalProcessedBytes, srcPtr, inputNbytes);
totalProcessedBytes += inputNbytes;

// Update checksum to prevent optimization
// Use the copied data to prevent compiler optimization
// Calculate checksum from both source and destination data
checksum ^= reinterpret_cast<uintptr_t>(srcPtr);
if (inputNbytes >= sizeof(uint64_t)) {
// XOR with first 8 bytes of copied data to ensure it's actually used
checksum ^= *reinterpret_cast<const uint64_t*>(
output.get() + totalProcessedBytes);
}

totalProcessedBytes += inputNbytes;
}
} else {
// Without prefetching - baseline implementation
Expand All @@ -358,10 +411,17 @@ int Rebatch::rebatch_single_thread() {

// Perform the memcpy operation
memcpy(output.get() + totalProcessedBytes, srcPtr, inputNbytes);
totalProcessedBytes += inputNbytes;

// Update checksum to prevent optimization
// Use the copied data to prevent compiler optimization
// Calculate checksum from both source and destination data
checksum ^= reinterpret_cast<uintptr_t>(srcPtr);
if (inputNbytes >= sizeof(uint64_t)) {
// XOR with first 8 bytes of copied data to ensure it's actually used
checksum ^= *reinterpret_cast<const uint64_t*>(
output.get() + totalProcessedBytes);
}

totalProcessedBytes += inputNbytes;
}
}

Expand All @@ -376,12 +436,8 @@ int Rebatch::rebatch_single_thread() {
*/
folly::coro::Task<int> Rebatch::rebatch_multi_thread(
std::shared_ptr<folly::Executor> pool) {
// Create a fresh batch of tensors for this operation
auto batch = std::make_shared<TensorBatch>(
tensors_per_batch_,
*size_distribution_,
output_tensor_size_,
*input_memory_pool_);
// Get a pregenerated batch for this operation (avoids on-the-fly generation)
auto batch = getNextBatch();

const auto& tensors = batch->getTensors();
const size_t numTensors = tensors.size();
Expand Down Expand Up @@ -429,10 +485,18 @@ folly::coro::Task<int> Rebatch::rebatch_multi_thread(
output->data() + output_offset + totalProcessedBytes,
srcPtr,
inputNbytes);
totalProcessedBytes += inputNbytes;

// Update checksum to prevent optimization
// Use the copied data to prevent compiler optimization
// Calculate checksum from both source and destination data
checksum ^= reinterpret_cast<uintptr_t>(srcPtr);
if (inputNbytes >= sizeof(uint64_t)) {
// XOR with first 8 bytes of copied data to ensure it's actually
// used
checksum ^= *reinterpret_cast<const uint64_t*>(
output->data() + output_offset + totalProcessedBytes);
}

totalProcessedBytes += inputNbytes;
}

return static_cast<int>(totalProcessedBytes ^ checksum);
Expand Down
21 changes: 19 additions & 2 deletions packages/adsim/src/cpp2/server/dwarfs/Rebatch.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -140,7 +140,8 @@ class Rebatch : public Kernel {
size_t memory_pool_size_gb,
int prefetch_dist,
std::string input_str,
std::string distribution_file);
std::string distribution_file,
size_t num_pregenerated_batches = 100);

std::string init(
std::shared_ptr<AdSimHandleObjs> h_objs,
Expand Down Expand Up @@ -173,14 +174,19 @@ class Rebatch : public Kernel {
}
std::string distribution_file = config_d["distribution_file"].asString();

size_t num_pregenerated_batches = config_d.count("num_pregenerated_batches")
? static_cast<size_t>(config_d["num_pregenerated_batches"].asInt())
: 100;

return std::make_shared<Rebatch>(
tensors_per_batch,
num_threads,
output_tensor_size,
memory_pool_size_gb,
prefetch_dist,
input_str,
distribution_file);
distribution_file,
num_pregenerated_batches);
}

private:
Expand All @@ -191,10 +197,21 @@ class Rebatch : public Kernel {
int64_t prefetch_dist_;
std::string input_str_;
std::string distribution_file_;
size_t num_pregenerated_batches_;

std::shared_ptr<SizeDistribution> size_distribution_;
std::shared_ptr<InputMemoryPool> input_memory_pool_;

// Pregenerated batches for performance optimization
std::vector<std::shared_ptr<TensorBatch>> pregenerated_batches_;
std::atomic<size_t> batch_index_{0};

// Pregenerate N batches to avoid on-the-fly generation
void pregenerateBatches();

// Get next pregenerated batch in round-robin fashion
std::shared_ptr<TensorBatch> getNextBatch();

// Single-threaded rebatch operation
int rebatch_single_thread();

Expand Down