feat: Optimize serializer decompress buffer for BufferInputStream #11836
Conversation
facebook-github-bot
added
the
CLA Signed
✅ Deploy Preview for meta-velox canceled.
|
jinchengchenghh
force-pushed
the
opt_compress
branch
from
0a803d5 to
7cfade1
Compare
jinchengchenghh
changed the title
jinchengchenghh
force-pushed
the
opt_compress
branch
from
7e78159 to
5dfe257
Compare
|
Can you help review again? Thanks! @Yuhta |
| auto newBuf = folly::IOBuf::wrapBuffer( | ||
| current_->buffer + current_->position, readBytes); | ||
| if (result) { | ||
| result->prev()->appendChain(std::move(newBuf)); |
There was a problem hiding this comment.
The order of buffers here is incorrect. Just do result->appendToChain(std::move(newBuf)).
There was a problem hiding this comment.
The order is same with IOBufOutputStream::getIOBuf(), appendToChain is also OK. I will change as you say.
| common::CompressionKind compressionKind, | ||
| memory::MemoryPool& pool) { | ||
| const auto codec = common::compressionKindToCodec(compressionKind); | ||
| if (dynamic_cast<BufferInputStream*>(source)) { |
There was a problem hiding this comment.
Just do the dynamic_cast once: if (const auto* bufferSource = dynamic_cast<BufferInputStream*>(source))
There was a problem hiding this comment.
I implement the readBytes return IOBuf in FileInputStream, so we don't need to do the cast, and other compression code can also benefit from this.
| auto byteStream = createStream(byteRanges); | ||
| auto bufferStream = dynamic_cast<BufferInputStream*>(byteStream.get()); | ||
| for (int offset = 0; offset < streamSize;) { | ||
| auto iobuf = bufferStream->readBytes(streamSize / 8); |
There was a problem hiding this comment.
We are not testing the case one IOBuf chain cross 2 buffers. Let's cover that as well.
jinchengchenghh
changed the title
|
Failed by this issue #11857 |
|
Can you help review again? Thanks! @Yuhta |
| current_->position += size; | ||
| return iobuf; | ||
| } | ||
| auto iobuf = folly::IOBuf::create(size); |
There was a problem hiding this comment.
This allocation should be avoided. Can we make a chain that takeOwnership() of the buffers instead?
We would also need to put pins on the buffers so that they are not overwritten while we have IOBuf on top of them. If this feels too complicated we can leave this out for next PR and just cast to BufferInputStream for now.
There was a problem hiding this comment.
I think it is complicated, and it breaks the FileInputStream should only reserve memory as QueryConfig::readBufferSize. Alternatively, we could add a new argument MemoryPool* pool = nullptr, allocated the Buffer from pool, and wrap the buffer as IOBuf. Which way do you think is better? This maybe in a follow up PR.
jinchengchenghh
changed the title
jinchengchenghh
force-pushed
the
opt_compress
branch
from
ce58686 to
38fe41c
Compare
|
I move to the cast to BufferInputStream as you say, and compress Row has merged, so I update it in same way. Can you help review again? Thanks~ @Yuhta |
|
Mac13 failed by arrow download, velox release uts failed by existing issue hdfs test, other tests passed. |
|
Can you help review again? Thanks! @Yuhta |
Read the BufferInputStream by new added function readBytes to avoid copying from the buffer cache. Extract the uncompress process to a new function can help release the compressed buffer in time.
FileInputStream reuse the read buffer, so we can get the IOBuf without copy only if the required bytes is in one buffer, which is not the common case, so I don't do the optimization for it.