Andrey Yarovoy created HDDS-15485:
-------------------------------------
Summary: Avoid ByteBuffer.wrap on ChunkBuffer.put(byte[]) path
Key: HDDS-15485
URL: https://issues.apache.org/jira/browse/HDDS-15485
Project: Apache Ozone
Issue Type: Bug
Reporter: Andrey Yarovoy
The default on {{ChunkBuffer}} is:
{code:java}
default ChunkBuffer put(byte[] b,int offset,int length) {
return put(ByteBuffer.wrap(b, offset, length));
}
{code}
So every {{{}put(byte[], offset, length){}}}:
# Allocates a new {{ByteBuffer}} view ({{{}HeapByteBuffer{}}} over the same
backing array).
# Dispatches to {{put(ByteBuffer)}}, which for list/incremental buffers runs a
segment loop that adjusts the wrapper’s position/limit and copies via
{{{}ByteBuffer.put(ByteBuffer){}}}.
On the client write path, {{BlockOutputStream.write(byte[], …)}} calls
{{currentBuffer.put(b, off, writeLen)}} once per slice that fits in the current
chunk buffer. Large writes can do that many times per {{write()}} call.
{{wrap}} itself is “cheap” (no array copy), but per call it still pays object
allocation + dispatch into the {{ByteBuffer}} put path. On a hot path called
for every slice of every {{{}write(byte[]){}}}, that cost is measurable under
load.
the fix would be micro-optimization, but should reduce memory allocations on
hot path and reduce number of objects for garbage collection.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
