This benchmark suite evaluates the throughput, latency, and resource efficiency of several Go-based communication frameworks under high-concurrency scenarios.
In this benchmarks 1,2,4,8,16 clients send echo requests to the server as follows:
- Each client executes all echo requests simultaneously in separate goroutines.
- The same data is used for all participants, it is generated once and then used by everyone.
- Size of the data varies from 17 to 1024 bytes.
- The size of the read and write buffers is limited to 4096 bytes (except Connect).
- The delay of each response on the server is 30 ms.
- The received data is checked - it must match the sent data.
Results were obtained on a single laptop (with the connected charger and fan running at full speed):
- CPU: AMD Ryzen 7 PRO 5850U with Radeon Graphics
- OS: Debian 6.1.148-1 x86_64 GNU/Linux
- Go: 1.24.1
using the following commands:
GEN_SIZE=400000 go test -bench BenchmarkQPS -count=10 -timeout=30m
go test -bench BenchmarkFixed -benchtime=100000x -benchmem -count=10Only the top 5 fastest cases are shown below. Full results are available in qps.csv.
Head-of-Line (HOL) blocking refers to the inability to send multiple concurrent requests over a single connection (i.e. you cannot send a new request until the one before it is finished). This behavior is characteristic of HTTP/1.1 and is precisely what HTTP/2 addresses via multiplexing.
The following participants are limited by HOL blocking:
- DRPC
In these benchmarks, they all will hit a "performance ceiling" of approximately 33 iterations per connection (qps.csv) due to the 30ms server delay:
To get more comparable results, let's check how well all participants can handle 100,000 simultaneous requests:
