cpitman / grpc-rest-bench

A microbenchmark of Java gRPC and Spring Rest

gRPC vs Spring Rest Microbenchmark

This project is an implementation of the simplest (in terms of implementation) possible microbenchmark that compares a single server synchronously echo'ing of data, with one request in flight at a time.

Both the client and server run in one process, communicating over localhost. That leaves the following areas to be measured by the microbenchmark:

• Serialization and Deserialization (JSON vs protobuf)
• Framework/Server Overhead (Spring Rest+Undertow vs gRPC)

Running the Benchmark

You can run this benchmark locally by cloning this repository, and then running mvn compile exec:exec.

This will start the servers and execute both benchmarks. Each benchmark is executed 10 times. Each time, it runs 20 warmup iterations, and then 20 iterations to capture results. At the end, aggregated results are printed, and the application exits.

Results

Running on my desktop (Windows 10, Intel i7-4790K), here were the results:

Benchmark                   Mode  Cnt     Score    Error  Units
GrpcBenchmark.callService  thrpt  200  3556.576 ± 38.107  ops/s
RestBenchmark.callService  thrpt  200  1810.138 ± 25.264  ops/s

Qualitatively, both benchmarks fully utilize a single CPU core (since only a single concurrent request is used, this is basically the cap). That implies both suites are CPU constrained for this test. That isn't surprising, since there is not real network between client and server.