Armeria server performance analaysis about http4s-armeria

[ikhoon]

Is it resolved?

I think there is a room we can optimize http4s-armeria performance. 💪

Loads of the benchmark were a simple "hello world" text message with HTTP/1.1.
Because of the simple load, the bottleneck point which mostly consumes CPU was a conversion from Reactive Streams to fs2 Stream and vise versa. fs2-reactive-streams is used for the conversion.

I only got rid of a conversion that converts from fs2 streams to Reactive Streams for HttpResponse.
Another remaining conversion is HttpRequest(Reactive Streams) to FS2 streams using fs2-reactive-streams.
I am going to remove fs2-reactive-streams dependency and implement an optimized version for fs2.

[ikhoon]

Agreed. Benchmark results might be changed somehow. I will run benchmarks soon.

[rossabaker]

@hamnis set up a server-example project, which we've used to run some crude benchmarks between the various backends. It doesn't show off much yet, but it gives us a baseline for pings.

[ngbinh]

Is it resolved? Congrats on the new release!

[ikhoon]

1st Benchmark:

• Environment: MacBook Pro, 2.2 GHz 6-Core Intel Core i7
• Benchmark code: hamnis/http4s-server-example@0cfeee3
• Result
  • http4s-armeria

    $ wrk -t12 -c400 -d30s http://127.0.0.1:8080/hello
    
    Running 30s test @ http://127.0.0.1:8080/hello
      12 threads and 400 connections
      Thread Stats   Avg      Stdev     Max   +/- Stdev
        Latency    11.76ms    2.97ms  86.66ms   89.53%
          Req/Sec     1.71k     0.94k    3.64k    52.50%
      611170 requests in 30.02s, 107.62MB read
      Socket errors: connect 155, read 170, write 0, timeout 0
    Requests/sec:  20358.15
    Transfer/sec:      3.58MB
    

  • Blaze

    $ wrk -t12 -c400 -d30s http://127.0.0.1:8080/hello
    Running 30s test @ http://127.0.0.1:8080/hello
      12 threads and 400 connections
      Thread Stats   Avg      Stdev     Max   +/- Stdev
        Latency     2.94ms    2.19ms 107.06ms   96.86%
        Req/Sec     6.99k     3.06k   14.33k    53.22%
      2506706 requests in 30.03s, 351.81MB read
      Socket errors: connect 155, read 277, write 0, timeout 0
    Requests/sec:  83481.35
    Transfer/sec:     11.72MB
    

  • Plain Armeria

    $ wrk -t12 -c400 -d30s http://127.0.0.1:8080/hello
    
    Running 30s test @ http://127.0.0.1:8080/hello
      12 threads and 400 connections
      Thread Stats   Avg      Stdev     Max   +/- Stdev
        Latency     3.16ms  450.83us  28.96ms   87.82%
        Req/Sec     6.37k     3.29k   13.52k    47.94%
      2280593 requests in 30.01s, 401.54MB read
      Socket errors: connect 155, read 142, write 0, timeout 0
    Requests/sec:  75992.98
    Transfer/sec:     13.38MB
    

• Bottleneck
  • http4s-armeria consumes all CPU cycles, it hits 100%
  • http4s-blaze hits: 50~60%
• Analysis
  • The performance(Requests/sec) of http4s-armeria was increased 3~4 times by removing toUnicastPublisher operation which converts Stream[F, HttpObject] to Publihser[HttpObject]
• Conclusion
  • We cannot release the initial version of http4s-armeria until solving the performance problem on the high CPU utilization.
  • We might need an obtimized converter for Reactive Streams?
  • ???

[ngbinh]

Thanks for the explanation. Looking forward to it!

[andyczerwonka]

Is it resolved?

I think there is a room we can optimize http4s-armeria performance. 💪

Loads of the benchmark were a simple "hello world" text message with HTTP/1.1. Because of the simple load, the bottleneck point which mostly consumes CPU was a conversion from Reactive Streams to fs2 Stream and vise versa. fs2-reactive-streams is used for the conversion.

I only got rid of a conversion that converts from fs2 streams to Reactive Streams for HttpResponse. Another remaining conversion is HttpRequest(Reactive Streams) to FS2 streams using fs2-reactive-streams. I am going to remove fs2-reactive-streams dependency and implement an optimized version for fs2.

@ikhoon was the optimization to fs2 done?

[ikhoon]

It was partially done and is still in progress. But there were many performance optimizations on the Armeria side.
I strongly believe that http4s-armeria can be used in production.

[danicheg]

@ikhoon It's been two years since you have done the benchmark. It'd be nice if you can run it with the new http4s-armeria/blaze. I believe some numbers must change in the benchmark results.