Use Netty's optimized UTF-8 encoding if available [SPR-17558]

[spring-projects-issues]

Mark Paluch opened SPR-17558 and commented

CharSequenceEncoder uses CharBuffer.wrap and charset.encode to encode String data to its binary representation. There can be two optimizations made here:

1. CharSequenceEncoder uses typically UTF-8 encoding and Java's UTF-8 encoder requires significant computing time. It would make sense to detect this case and whether netty is on the class path to use netty's optimized UTF-8 encoding via ByteBufUtil.writeUtf8(…)
2. Encoding creates a new unpooled ByteBuffer when calling Charset.encode. Netty's ByteBufUtil.encodeString() can encode a String to a pooled buffer that reduces GC pressure.

See also attached profiling snapshot.

---

Attachments:

• LettucetestApplication-2018-12-03-RenderStringToBytesCharsetEncoder.snapshot.zip (2.40 MB)
• Write as String.png (431.36 kB)

Referenced from: commits 5a8b8b1, 6361b0c, a00be62, 4955d08

0 votes, 5 watchers

[spring-projects-issues]

Arjen Poutsma commented

It would make sense to detect this case and whether netty is on the class path.

It would not make sense, unfortunately: whether netty is on the class path is not a valid trigger. Users can choose to use a different DataBuffer implementation, or—when using the webclient on Tomcat, Jetty, or Undertow—ByteBuf will be on the classpath, but should only be used for the client-side; not server-side.

We could check whether a data buffer is an instance of NettyDataBuffer, and then call getNativeBuffer() to get access to the ByteBuf.

However, I think that's a bit ugly, so I'd rather expose the necessary functionality through the DataBuffer and DataBufferFactory abstractions, delegate to Netty's ByteBufUtil for NettyDataBuffer, and use a different, less optimal implementation for the DefaultDataBuffer.

[spring-projects-issues]

Brian Clozel commented

Hey Arjen, that’s what I had in mind as well.
I can work on that and ask for a review or feel free to reassign this issue to yourself. I self-assigned it initially to make sure to benchmark the fix.