[Twisted-Python] Benchmark of Python WSGI servers

[Michael Thompson]

On 19 March 2011 08:29, Glyph Lefkowitz <glyph at twistedmatrix.com> wrote:
> On Mar 18, 2011, at 9:35 PM, exarkun at twistedmatrix.com wrote:
>
> A multi-process WSGI container might actually be of practical use, since
> it may make more cores available to your server.  If an application is
> bottlenecked on CPU rather than some high-latency operation (as you can
> only process as many concurrent requests as you have threads in your
> threadpool), more cores can help.
>
> +1.  This will actually speed up your application code, not just the server
> :).

I agree that threads and processes is going to be the first step to
scaling but I don't think there's much advantage putting that in
twisted given that there are other solutions for this.

I was hoping that changing the reactor would give an easy way to make
Twisted look better, from a PR point of view. To check I ran a
benchmark on two EC2 m1.small machines using

httperf --timeout=5 --client=0/1 --server=ip-10-234-217-116
--port=8443 --uri=/ --rate=1000 --send-buffer=4096 --recv-buffer=16384
--num-conns=4000 --num-calls=10

gevent 1998 req/s.
select  734 req/s
poll      687 req/s
epoll    694 req/s

So looks like the reactor has nothing to do with it. Here some profile
results for the select reactor.

         2716620 function calls (2716586 primitive calls) in 21.634 CPU seconds

   Ordered by: internal time
   List reduced from 340 to 34 due to restriction <0.10000000000000001>

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
       55    5.994    0.109    6.141    0.112 {select.select}
    22871    0.992    0.000    0.992    0.000 {built-in method acquire}
     5706    0.642    0.000    1.927    0.000 wsgi.py:168(__init__)
   165474    0.521    0.000    0.528    0.000 http.py:626(__setattr__)
     5706    0.394    0.000    1.778    0.000 http.py:922(write)
     5706    0.379    0.000    7.879    0.001 basic.py:543(dataReceived)
   142650    0.363    0.000    0.363    0.000 {method 'lower' of 'str' objects}
    34236    0.341    0.000    0.582    0.000
http_headers.py:13(_dashCapitalize)
    74178    0.324    0.000    0.712    0.000 http_headers.py:222(getRawHeaders)
    34236    0.319    0.000    0.954    0.000
http_headers.py:249(_canonicalNameCaps)
    22824    0.305    0.000    7.434    0.000 http.py:1537(lineReceived)
   133660    0.283    0.000    0.283    0.000 {method 'get' of 'dict' objects}
     5700    0.261    0.000    0.494    0.000 abstract.py:186(doWrite)
   108414    0.256    0.000    0.256    0.000 {method 'split' of 'str' objects}
     5706    0.252    0.000    0.739    0.000 wsgi.py:273(_sendResponseHeaders)
    11412    0.242    0.000    0.527    0.000 http.py:1589(headerReceived)
     5706    0.242    0.000    0.601    0.000 http.py:606(__init__)
    12043    0.242    0.000    9.162    0.001
selectreactor.py:144(_doReadOrWrite)
    39942    0.237    0.000    0.357    0.000 http_headers.py:186(setRawHeaders)
     5706    0.219    0.000    0.235    0.000 http.py:217(datetimeToString)
     5706    0.200    0.000    4.800    0.001 server.py:109(process)
    28530    0.197    0.000    0.307    0.000 http.py:1773(_escape)
     6301    0.196    0.000    0.240    0.000 base.py:688(callLater)
    12043    0.194    0.000    9.432    0.001 context.py:32(callWithContext)
   125616    0.194    0.000    0.194    0.000 {len}
     5706    0.193    0.000    0.526    0.000 http.py:1701(requestDone)
    91366    0.193    0.000    0.193    0.000 {isinstance}
    22824    0.173    0.000    0.204    0.000 base.py:97(reset)
    97006    0.169    0.000    0.169    0.000 {method 'replace' of
'str' objects}
    12043    0.169    0.000   10.085    0.001 log.py:71(callWithLogger)
    45648    0.168    0.000    1.207    0.000
http_headers.py:239(getAllRawHeaders)
    22856    0.148    0.000    0.148    0.000 {built-in method release}
5715/5712    0.147    0.000    0.308    0.000 defer.py:467(_runCallbacks)
    28530    0.144    0.000    0.418    0.000 http.py:856(getHeader)