Application-Level Service Discovery Benchmark Testing

Mon, 01 Jan 0001 00:00:00 +0000

1 Benchmark Conclusions

Compared to version 2.x, Dubbo 3 version

Resource utilization for service discovery has significantly improved.
- Compared to interface-level service discovery, the resident memory on a single machine decreased by 50%, and the GC consumption during address changes decreased an order of magnitude (hundreds -> tens).
- Compared to application-level service discovery, the resident memory on a single machine decreased by 75%, and the number of GCs approached zero.

Here are the detailed pressure test processes and data

Mon, 01 Jan 0001 00:00:00 +0000

The implementation of the Dubbo protocol in Dubbo3 is basically on par with the Dubbo2 version in terms of performance.
Since the Triple protocol is built on HTTP/2, RPC calls over a single link do not show an improvement compared to Dubbo2 based on TCP, and there is actually a certain decrease in some calling scenarios. However, the Triple protocol has greater advantages in gateway penetration, universality, and overall throughput improvement brought by the Stream communication model.
Triple is expected to perform better under gateway proxy scenarios, and benchmarks were not provided in the current stress test environment.

	Description
Machine	4C8G Linux JDK 1.8 (Provider) 4C8G Linux JDK 1.8 (Consumer)
Stress Test Cases	RPC method types include: no-parameter no return value, ordinary pojo return value, pojo list return value Version 2.7 Dubbo protocol (Hessian2 serialization) Version 3.0 Dubbo protocol (Hessian2 serialization) Version 3.0 Dubbo protocol (Protobuf serialization) Version 3.0 Triple protocol (Protobuf serialization) Version 3.0 Triple protocol (Protobuf plus Hessian2 serialization)
Stress Testing Method	In a single link scenario, the consumer starts 32 concurrent threads (the current machine configuration has a balanced qps rt concurrency), continuously collecting test data Test data is obtained from https://github.com/apache/dubbo-benchmark

	Dubbo + Hessian2 2.7	Dubbo + Hessian2 3.0	Dubbo + Protobuf 3.0	Triple + Protobuf 3.0	Triple + Protobuf(Hessian) 3.0
No-parameter Method	30333 ops/s 2.5ms P99	30414 ops/s 2.4ms P99	24123 ops/s 3.2ms P99	7016 ops/s 8.7ms P99	6635 ops/s 9.1ms P99
Pojo Return Value	8984 ops/s 6.1 ms P99	12279 ops/s 5.7 ms P99	21479 ops/s 3.0 ms P99	6255 ops/s 8.9 ms P99	6491 ops/s 10 ms P99
Pojo List Return Value	1916 ops/s 34 ms P99	2037 ops/s 34 ms P99	12722 ops/s 7.7 ms P99	6920 ops/s 9.6 ms P99	2833 ops/s 27 ms P99

Mon, 01 Jan 0001 00:00:00 +0000

dubbo-benchmark