ASGI support

[remittor]

Somehow I made support for ASGI: https://github.com/remittor-pr/fastwsgi/commits/devrem
It was very difficult. Some points still need to be improved.

Test app: server.py , asgi_srv.py

WSGI

$ python3 server.py -g fw -v 2

$ wrk -t1 -c1 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 1 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   138.25us  369.95us   8.61ms   98.67%
    Req/Sec     9.35k   505.91    10.31k    86.09%
  140470 requests in 15.10s, 20.90MB read
Requests/sec:   9302.95
Transfer/sec:      1.38MB

$ wrk -t1 -c100 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     0.98ms  383.96us   9.54ms   72.26%
    Req/Sec   100.19k    10.81k  108.90k    90.67%
  1494436 requests in 15.00s, 222.33MB read
Requests/sec:  99619.27
Transfer/sec:     14.82MB

ASGI

$ python3 asgi_srv.py -g fw -v 2

$ wrk -t1 -c1 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 1 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     1.49ms    2.47ms  31.42ms   82.87%
    Req/Sec     4.68k     4.47k    9.84k    30.00%
  69851 requests in 15.01s, 10.39MB read
Requests/sec:   4653.26
Transfer/sec:    708.90KB

$ wrk -t1 -c100 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     2.60ms  290.70us   7.67ms   85.20%
    Req/Sec    38.47k     2.63k   41.98k    76.00%
  574024 requests in 15.00s, 85.40MB read
Requests/sec:  38264.41
Transfer/sec:      5.69MB

[remittor]

For comparison

Socketify v0.0.20 (ASGI)

$ python3 asgi_srv.py -g si -v 2

$ wrk -t1 -c1 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 1 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     1.49ms    2.39ms  10.13ms   82.60%
    Req/Sec     3.86k     3.51k    8.06k    30.00%
  57623 requests in 15.00s, 8.63MB read
Requests/sec:   3841.26
Transfer/sec:    588.94KB

$ wrk -t1 -c100 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     4.02ms  641.52us  13.43ms   95.51%
    Req/Sec    24.96k     0.95k   27.53k    85.33%
  372516 requests in 15.00s, 55.78MB read
Requests/sec:  24832.23
Transfer/sec:      3.72MB

UviCorn v0.21.1 (ASGI)

$ python3 asgi_srv.py -g uc -v 2

$ wrk -t1 -c1 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 1 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   371.36us  110.32us   6.02ms   98.63%
    Req/Sec     2.72k   100.59     2.90k    65.33%
  40541 requests in 15.00s, 6.50MB read
Requests/sec:   2702.70
Transfer/sec:    443.41KB

$ wrk -t1 -c100 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    21.77ms    6.74ms  45.35ms   62.39%
    Req/Sec     4.60k   322.90     5.27k    64.67%
  68637 requests in 15.01s, 11.00MB read
Requests/sec:   4572.99
Transfer/sec:    750.26KB

[remittor]

Developer @cirospaciari writes here: emmett-framework/granian#30
I noticed 2 days ago that callback task.done_cb() (add_done_callback) is taking control too late. But I didn't attach any importance to it.

Now I tried to completely remove this callback (it turned out to be completely unnecessary).
Speed up:

$ wrk -t1 -c100 -d15 http://10.10.2.119:5000/plaintext
Running 15s test @ http://10.10.2.119:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     2.19ms  389.22us   8.62ms   75.86%
    Req/Sec    45.26k     3.51k   47.71k    88.67%
  675536 requests in 15.00s, 100.50MB read
Requests/sec:  45024.62
Transfer/sec:      6.70MB

[cirospaciari]

AWESOME maybe renaming to fastXSGI or something like this will be more intuitive, this will be the fastest CPython WSGI and ASGI.
Kudos @remittor
I will try to improve socketify to be able to compete :D

fastWSGI in CPython scored 1.3mi req/s
Granian WSGI in CPython scored 800k req/s
socketify WSGI in CPython scored 700k req/s
socketify WSGI in PyPy scored 1.5mi req/s
I think I can get it to the level of fastWSGI if I implement it on native C, and maybe 2 to 3 mi req/s using PyPy.

Python future is awesome! I think this will be the end of uvicorn and meinheld without performance competition era.

[remittor]

I managed to make support for the Quart framework: https://github.com/remittor-pr/fastwsgi/commits/devrem

Test app: asgi_srv.py

Quart + Hypercorn (default):

$ python3 asgi_srv.py -g hc -q

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    87.65ms   20.93ms 138.38ms   64.27%
    Req/Sec     1.14k    97.32     1.65k    76.67%
  17072 requests in 15.01s, 2.22MB read
Requests/sec:   1137.62
Transfer/sec:    151.21KB

Quart + Uvicorn:

$ python3 asgi_srv.py -g uc -q

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    51.25ms    7.03ms  68.33ms   64.60%
    Req/Sec     1.96k   167.61     2.22k    66.67%
  29199 requests in 15.01s, 3.68MB read
Requests/sec:   1945.69
Transfer/sec:    250.81KB

Quart + FastWSGI:

$ python3 asgi_srv.py -g fw -q

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    35.54ms   21.75ms  87.06ms   52.51%
    Req/Sec     2.83k   232.65     3.41k    80.00%
  42324 requests in 15.01s, 7.07MB read
Requests/sec:   2820.07
Transfer/sec:    482.25KB

[cirospaciari]

Great! 🚀 Quart it self is the only ASGI that performs worse on socketify CPython for some reason.

[remittor]

Quart + Socketify v0.0.20

$ python3 asgi_srv.py -g si -q

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    44.41ms    5.14ms  60.37ms   88.44%
    Req/Sec     2.26k   159.60     2.44k    74.00%
  33727 requests in 15.00s, 5.05MB read
Requests/sec:   2248.19
Transfer/sec:    344.69KB

[cirospaciari]

interesting on your hardware is faster than uvicorn, but is still slow.

[remittor]

FastAPI + Uvicorn

$ python3 asgi_srv.py -g uc -f

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    26.69ms    8.92ms  52.11ms   55.78%
    Req/Sec     3.75k   302.44     4.35k    61.33%
  56066 requests in 15.01s, 7.86MB read
Requests/sec:   3735.34
Transfer/sec:    536.22KB

FastAPI + FastWSGI

$ python3 asgi_srv.py -g fw -f -v 2

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    10.30ms    1.69ms  19.97ms   93.81%
    Req/Sec     9.75k   415.64    10.56k    65.33%
  145591 requests in 15.00s, 21.38MB read
Requests/sec:   9703.95
Transfer/sec:      1.43MB

[remittor]

BlackSheep+ Uvicorn

$ python3 asgi_srv.py -g uc -b

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     4.73ms    1.64ms  12.64ms   61.88%
    Req/Sec    20.99k   522.65    22.10k    77.33%
  313343 requests in 15.01s, 39.45MB read
Requests/sec:  20881.40
Transfer/sec:      2.63MB

BlackSheep+ FastWSGI

$ python3 asgi_srv.py -g fw -b -v 2

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:5000/plaintext
Running 15s test @ http://127.0.0.1:5000/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     3.71ms  786.68us  13.64ms   96.72%
    Req/Sec    27.29k     0.98k   29.29k    64.00%
  407352 requests in 15.00s, 54.00MB read
Requests/sec:  27148.81
Transfer/sec:      3.60MB

[jamesroberts]

Apologies, I have been away for a while.

This looks awesome!
I will have a look at the PR when I get a chance later today.
Great work @remittor!