Understanding Dubbo’s Network Transport Layer and Hessian2 Serialization

1. Dubbo Network Transport Layer

Dubbo’s transport layer depends on the serialization layer and handles one‑way message transmission. When a consumer calls a provider, method names and parameters are sent as binary data, decoded by a unified protocol, allowing upper‑level models to work with objects without worrying about transport details.

The transport layer supports multiple protocols (Dubbo, RMI, Hessian, WebService, HTTP) and serialization methods (Hessian2, dubbo, JSON, Java native).

2. Dubbo Protocol

Dubbo uses its own protocol with a fixed‑length header and variable‑length body; the header records serialization type, request status, and data length, while the body contains the serialized request/response object.

When Netty is the transport framework, the client and server correspond to NettyClient and NettyServer. The default codec is DubboCountCodec, whose inner DubboCodec implements the protocol’s encode/decode logic.

During a request, the consumer’s parameters are encoded by InternalEncoder before being sent.

On the provider side, InternalDecoder assembles possibly fragmented packets into a complete Dubbo protocol message using an internal ChannelBuffer, then decodes it.

Because Netty may deliver incomplete or multiple Dubbo packets, InternalDecoder uses a buffer to handle both scenarios, rolling back the read index when more input is needed.

3. Hessian2 Serialization Protocol

Hessian2, an open‑source binary serialization protocol from Caucho, is used by Dubbo via the com.alibaba.caucho.hessian.io package. After the Dubbo header is processed, objects are serialized/deserialized with Hessian2Output and Hessian2Input.

3.1 String Serialization and Deserialization

Strings dominate the binary payload. Hessian2 encodes characters in UTF‑8, splitting long strings into chunks marked by ‘S’ (final chunk) or ‘R’ (non‑final chunk), each followed by a 16‑bit length indicating character count.

During serialization, Hessian2Output.printString iterates over characters, writes encoded bytes to a buffer; during deserialization, Hessian2Input.readString parses the marker, reads the length, and reconstructs characters via parseChar / parseUTF8Char.

4. Optimization Points

4.1 Reduce Array Copies

Dubbo’s InternalDecoder copies the packet into a ChannelBuffer, which is then copied again by HessianInput. Removing the internal 256‑byte cache in HessianInput and deserializing directly from the ChannelBuffer eliminates redundant copies.

4.2 Use Netty 4 ByteToMessageDecoder

Netty 4’s ByteToMessageDecoder provides a reusable cumulation buffer, reducing memory copies and simplifying decoder code; Dubbo’s decoder can be refactored to extend this class for better performance.

4.3 Compress Binary Data

Hessian2 supports compression. In a test, raw Java serialization produced 10 320 bytes, Hessian2 6 844 bytes, and compressed Hessian2 3 915 bytes (57 % of uncompressed). Compression roughly doubles serialization time, so it should be used only when network bandwidth is a bottleneck.

4.4 Choose Alternative Serialization

Newer serializers such as Kryo, FST (Java‑only) and cross‑language formats like ProtoBuf or Thrift outperform Hessian2. Projects like Dubbox already replace Hessian2 with Kryo/FST to boost Dubbo performance.

5. References

Dubbo‑2.5.4 – GitHub Dubbo Protocol Reference – dubbo.io Hessian Serialization – caucho.com Dubbox Serialization Benchmarks – dangdangdotcom.github.io