[VL] Support GPU async native shuffle read#12370
Open
marin-ma wants to merge 9 commits into
Open
Conversation
|
Run Gluten Clickhouse CI on x86 |
marin-ma
force-pushed
the
gpu-async-native-shuffle-read
branch
from
e5cc3aa to
0713a26
Compare
|
Run Gluten Clickhouse CI on x86 |
marin-ma
force-pushed
the
gpu-async-native-shuffle-read
branch
from
0713a26 to
bebb887
Compare
|
Run Gluten Clickhouse CI on x86 |
marin-ma
force-pushed
the
gpu-async-native-shuffle-read
branch
from
bebb887 to
5ecc27c
Compare
|
Run Gluten Clickhouse CI on x86 |
2 similar comments
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
marin-ma
force-pushed
the
gpu-async-native-shuffle-read
branch
from
f840d7a to
c1d2691
Compare
|
Run Gluten Clickhouse CI on x86 |
2 similar comments
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
marin-ma
force-pushed
the
gpu-async-native-shuffle-read
branch
from
b716f71 to
7c4a796
Compare
|
Run Gluten Clickhouse CI on x86 |
9 similar comments
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
marin-ma
force-pushed
the
gpu-async-native-shuffle-read
branch
from
ad69c7d to
096215f
Compare
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
|
Run Gluten Clickhouse CI on x86 |
marin-ma
force-pushed
the
gpu-async-native-shuffle-read
branch
from
1c29a29 to
33e5b75
Compare
|
Run Gluten Clickhouse CI on x86 |
|
|
||
| private def assertPositiveBlockSize(blockId: BlockId, blockSize: Long): Unit = { | ||
| if (blockSize < 0) { | ||
| throw BlockException(blockId, "Negative block size " + size) |
Comment on lines
+61
to
+62
| LOG(INFO) << "Trying to get from cached buffer queue. Queue length: " << queue_.size() | ||
| << ", total size in queue: " << totalSize_ << ", current batch size: " << batch->numBytes() << std::endl; |
|
|
||
| notFull_.wait(lock, [&]() { return noMoreBatches_ || totalSize_ + batchSize <= capacity_; }); | ||
| if (noMoreBatches_) { | ||
| LOG(WARNING) << "Discard batch due to calling put() after noMorBatches()."; |
Comment on lines
225
to
229
| // Stop reading more streams. Blocked by the native reader threads. | ||
| jniWrapper.stop(shuffleReaderHandle) | ||
| onComplete.foreach(_()) | ||
| // Would remove the resource object from registry to lower GC pressure. | ||
| TaskResources.releaseResource(resourceId) |
|
Run Gluten Clickhouse CI on x86 |
marin-ma
changed the title
|
Run Gluten Clickhouse CI on x86 |
|
|
||
| private def assertPositiveBlockSize(blockId: BlockId, blockSize: Long): Unit = { | ||
| if (blockSize < 0) { | ||
| throw BlockException(blockId, "Negative block size " + size) |
Comment on lines
+304
to
+314
| ReaderThreadPool* VeloxBackend::getReaderThreadPool() { | ||
| static std::once_flag readerThreadPoolInit; | ||
| std::call_once(readerThreadPoolInit, [this] { | ||
| const auto configuredThreads = | ||
| backendConf_->get<int32_t>(kShuffleReaderThreads, static_cast<int32_t>(std::thread::hardware_concurrency())); | ||
| // std::thread::hardware_concurrency() can return 0; | ||
| const auto numThreads = configuredThreads > 0 ? configuredThreads : 1; | ||
| readerThreadPool_ = std::make_unique<ReaderThreadPool>(numThreads); | ||
| }); | ||
| return readerThreadPool_.get(); | ||
| } |
Comment on lines
+40
to
+43
| for (auto& task : tasks) { | ||
| tasks_.push({std::move(task), priority}); | ||
| } | ||
| } |
Comment on lines
+85
to
+94
| auto& prioritizedTask = tasks_.top(); | ||
| LOG(INFO) << "Worker thread " << std::this_thread::get_id() << " is executing a task with priority " | ||
| << prioritizedTask.priority; | ||
| task = std::move(prioritizedTask.task); | ||
| tasks_.pop(); | ||
| } | ||
|
|
||
| if (task) { | ||
| task(); | ||
| } |
Comment on lines
+132
to
134
| void VeloxGpuHashShuffleReaderDeserializer::read() { | ||
| std::shared_ptr<arrow::io::InputStream> inputStream = nullptr; | ||
|
|
Comment on lines
+191
to
+200
| // Close input stream if it's still open. | ||
| if (inputStream != nullptr) { | ||
| GLUTEN_THROW_NOT_OK(inputStream->Close()); | ||
| } | ||
|
|
||
| // Decrement active reader count. | ||
| if (activeReaders_.fetch_sub(1, std::memory_order_acq_rel) == 1) { | ||
| batchQueue_->noMoreBatches(); | ||
| completionCV_.notify_all(); | ||
| } |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The parallelism of gpu stages is limited by the GPU concurrency, but the shuffle read process, which includes data fetching, decompression and deserialisation, are still running on CPU. In this case we can parallelise these process to produce the output data asynchronously.
This PR adopts a producer–consumer design. Producer threads asynchronously perform shuffle reads, including data fetching, decompression, and deserialization, and produce decoded data. The consumer (the main thread) retrieves the prepared data as it becomes available and creates the corresponding device buffers.