[Flask] uWSGI and Flask - Slow and varying response times

[Unai Rodriguez]

Fantastic collaborative effort!!! :) 

-- unai

On Thu, Jul 7, 2016, at 02:42 PM, Ziirish wrote:
> Hello,
> 
> I think you have your answer right with the last piece of information you
> gave us.
> 
> Your CPU clock is two times faster than your servers one.
> 
> Now your code is probably not multi threaded, it means your server will
> perform at least two times slower than your development machine to
> execute the same piece of code.
> 
> The x3 you are measuring is probably due to some context switch or
> something (your kernel is probably spitting the task in several pieces to
> let other processes perform well so these pieces of processes then need
> to be synced or something)
> 
> In conclusion, the delay you are observing is due to a lower CPU clock on
> the server. So unless you can multi thread your application there's not
> much to do.
> In the other hand, because you have a lot of cores, your server will be
> able to serve (at least) 8 simultaneous clients within the same time
> range.
> 
> You can also have a look at PyPy as already suggested to try to optimize
> the generated "bytecode". It can probably save you a few CPU cycles and
> thus decrease your whole compute time.
> 
> 
> 
> 
> Le 7 juillet 2016 05:40:30 GMT+02:00, Tim van der Linden <tim at shisaa.jp>
> a écrit :
> >On Wed, 06 Jul 2016 18:28:15 +0800
> >Unai Rodriguez <unai at sysbible.org> wrote:
> >
> >> Hi Tim,
> >
> >Hi Unai
> >
> >> For uwsgi/Python you could do the same, build a simple page and a
> >more
> >> complicated one and compare (no database queries to isolate).
> >
> >First a simple page on uWSGI using Flask and Jinja. The template only
> >consists out of the bare minimals for a HTML page and a single string
> >variable to be printed and the whole application is a single Python
> >file with only one view. The timings are done using the Werkzeug
> >profiler:
> >
> >Local: 7 ms for initial request, 0.5ms for all following requests
> >Server: 160 ms for initial request, 1ms for all following requests
> >
> >I will count that as identical. The Snakeviz graphs also look identical
> >for both.
> >
> >Then I added a few range(1, 1000) loops to be printed in Jinja2 and a
> >load of simple text:
> >
> >Local: 9 ms for initial request, 3 ms for all following requests
> >Server: 24 ms for initial request, 10 ms for all following requests
> >
> >Slowly a difference between the two start to form.
> >
> >I added some more elements to the view and template such as a WTF form,
> >but the timings did not went up much. The difference stayed however:
> >server being three times slower then local.
> >
> >I also added a couple of (fast) database queries in the mix, opening
> >and closing both cursors and connections per query:
> >
> >Local: 16 ms for initial request, 11 ms for all following requests
> >Server: 51 ms for initial request, 33 ms for all following requests
> >
> >> The server/os will have to be looked at some point as well. 
> >
> >This was interesting but maybe less important since this is testing
> >pure CPU throughput:, I yanked a primes Python script of the interwebs
> >to test the raw processing speed and ran it on both machines.
> >
> >Local: 3.3879199028 s
> >Server: 9.26460504532 s
> >
> >This is a script without any extras, one Python function tested with
> >Timeit and ran directly via the Python CLI. My local machine is three
> >times faster than the server.
> >
> >Installing Sysbench and doing a primers test (20000) outside of Python:
> >
> >Local: 18.1605 seconds
> >Server: 42.1942 seconds
> >
> >CPU info:
> >
> >Local: 1 x i74790K (4 Ghz)
> >Server: 8 x Xeon E7-4820 v2 (2 Ghz)
> >
> >> -- unai
> >
> >Cheers,
> >Tim
> >_______________________________________________
> >Flask mailing list
> >Flask at python.org
> >https://mail.python.org/mailman/listinfo/flask
>