🤗 Help me understand the performance

[Abhijeet104]

Default net/http gives me 16k TPS with 100% success, but fiber fails(~98% success) at 10k TPS for normal Hello World API Calls. How some of the benchmarks show it as better than http/net? 🤔

Code snippet Optional

func helloHandler(w http.ResponseWriter, r *http.Request) {
    if r.URL.Path != "/hello" {
        http.Error(w, "404 not found.", http.StatusNotFound) 
        return 
    }

    if r.Method != "GET" {
        http.Error(w, "Method is not supported.", http.StatusNotFound)
        return
    }
    fmt.Fprintf(w, "Hello, World 👋!")
}

func main() {
    http.HandleFunc("/hello", helloHandler)
    if err := http.ListenAndServe(":9999", nil); err != nil {
        log.Fatal(err)
    } 
}

Tested above code with sample code from fiber's git read me section. Tested using vegeta for 1m, with above rate, on same machine MAC book pro(16GB RAM, 12 cores)

[welcome[bot]]

Thanks for opening your first issue here! 🎉 Be sure to follow the issue template! If you need help or want to chat with us, join us on Discord https://gofiber.io/discord

[Abhijeet104]

Was unable to open discord site. so had to post it here.

[jxsl13]

can confirm, the invite seems to have expired.

[Fenny]

Welcome @Abhijeet104!

Both are relatively fast/equal when it comes to local benchmarking because you will hit your OS limit relatively soon. You have to watch out that your benchmarking tool does not use more CPU than your Go process.

The biggest difference between Fiber and net/http is the memory allocation, Fasthttp uses a worker pool with efficient parsing methods where net/http spawns a new go routine and allocates memory on each connection.

This difference is clearly visible with the techempower benchmarks where they use a Switched 10-gigabit ethernet pipe on a separate machine for benchmarking to replicate a real-world environment.

64 conns

fiber   : TotalAlloc = 2 MiB      Sys = 12 MiB    NumGC = 0
net/http: TotalAlloc = 1599 MiB   Sys = 17 MiB    NumGC = 229

// fiber
Running 10s test @ http://192.168.1.131:3000
  6 threads and 64 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   424.00us  160.78us  12.82ms   82.37%
    Req/Sec    23.03k   825.05    27.99k    72.73%
  1386015 requests in 10.10s, 202.24MB read
Requests/sec: 137229.86
Transfer/sec:     20.02MB

// net/http
Running 10s test @ http://192.168.1.131:3001
  6 threads and 64 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   515.06us  302.11us  10.19ms   88.70%
    Req/Sec    19.95k     0.87k   23.85k    70.91%
  1200295 requests in 10.10s, 154.53MB read
Requests/sec: 118842.31
Transfer/sec:     15.30MB

2500 conns

fiber   : TotalAlloc = 13 MiB     Sys = 30 MiB    NumGC = 3
net/http: TotalAlloc = 2652 MiB   Sys = 46 MiB    NumGC = 829

// fiber
Running 20s test @ http://192.168.1.131:3000
  6 threads and 2500 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     7.84ms   22.06ms 989.74ms   98.33%
    Req/Sec    33.61k    16.54k   66.85k    59.90%
  3344237 requests in 20.04s, 440.13MB read
Requests/sec: 166855.80
Transfer/sec:     21.96MB

// net/http
Running 20s test @ http://192.168.1.131:3001
  6 threads and 2500 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     8.25ms    5.20ms 330.37ms   95.68%
    Req/Sec    20.76k     8.05k   47.92k    44.25%
  2478725 requests in 20.01s, 347.49MB read
  Socket errors: connect 1481, read 0, write 0, timeout 0
Requests/sec: 123852.68
Transfer/sec:     17.36MB

So in a real-world scenario, your net/http server could run out of memory on high loads and be throttled. Where Fiber will keep serving requests because of the low memory footprint, this is exactly where Fasthttp was designed for.

PS: our discord invite link seems to be valid, could you try https://discord.gg/bSnH7db?

[nyaa8]

Fiber:

> bombardier -c 245 127.0.0.1:3000
Bombarding http://127.0.0.1:3000 for 10s using 245 connection(s)
[==========================================================================] 10s
Done!
Statistics        Avg      Stdev        Max
  Reqs/sec    132211.02    8799.84  155085.90
  Latency        1.85ms   152.64us    16.53ms
  HTTP codes:
    1xx - 0, 2xx - 1322000, 3xx - 0, 4xx - 0, 5xx - 0
    others - 0
  Throughput:    24.84MB/s

2020/10/04 22:40:59 Alloc = 3 MiB    TotalAlloc = 3 MiB    Sys = 69 MiB    NumGC = 0

net/http

> bombardier -c 245 127.0.0.1:3002
Bombarding http://127.0.0.1:3002 for 10s using 245 connection(s)
[==========================================================================] 10s
Done!
Statistics        Avg      Stdev        Max
  Reqs/sec    119381.83   17016.99  153692.31
  Latency        2.05ms   344.36us    36.02ms
  HTTP codes:
    1xx - 0, 2xx - 1193836, 3xx - 0, 4xx - 0, 5xx - 0
    others - 0
  Throughput:    22.43MB/s

2020/10/04 22:42:04 Alloc = 5 MiB    TotalAlloc = 1602 MiB    Sys = 70 MiB    NumGC = 542

🙇🏻‍♀️

Specifications

**OS** macOS Catalina (hackintosh) **Go** version go1.14.4 darwin/amd64 **RAM** 16 GB DDR3 **CPU** Intel i7 4790k ~ 4.40 Ghz