Improvements observed with larger buffer sizes and chunked encoding

[nairiti-sharma]

We observed 6X response time degradation on an API that fetches >7KB response with chunked encoding.

• Helidon Version: 4.0.11
• Helidon SE or Helidon MP : SE
• JDK version: GraalVM on JDK21
• OS: Linux
• Docker version (if applicable): NA

---

Problem Description

On H3, there was no setting specific to byte buffer and it was left at the default value (https://github.com/netty/netty/blob/netty-4.1.108.Final/transport/src/main/java/io/netty/channel/socket/nio/NioSocketChannel.java )

On H4, there is a default of 512B which is causing the performance degradation.

In the JFR we could see that the number of Socket Writes with H4 are 2 times more than H3, even when throughput of H3 is 4 times more than H4.

When we increased this value to 8KB (-Dserver.write-buffer-size=8192), performance is at par with H3.

H3 : RT - 4.02ms
H4 (Default) : RT - 24ms
H4 with write-buffer-size - 1KB : RT - 5.37ms
H4 with write-buffer-size - 8KB : RT - 4.23ms

Steps to reproduce

Any response that has size greater than 7KB and "Transfer-Encoding: chunked" , would reproduce this issue.

[spericas]

@nairiti-sharma Let's consider a buffer size of 2-4K as a potential new default. Apps can always increase this if necessary.

[nairiti-sharma]

These are the final numbers
H3 : RT - 3.87ms
H4 (Default) : RT - 28.5ms
H4 with write-buffer-size - 1KB : RT - 5.37ms
H4 with write-buffer-size - 4KB : RT - 5.24ms
H4 with write-buffer-size - 8KB : RT - 3.94ms

Setting 4K by default should work well. Services can tune this further as per their requirement.