Hello Roberto,

that strange but I got the same results. I checked twicely that
patches correctly applied and compiled, but I don't see any difference
in benchmark results? looks like bottleneck is somewhere else.

the main problem that if we are going to keep active 100-200K
connections - this can be an issue, but maybe we are already at some
top limit, so maybe it's ok - I'am not sure.

also, the main my question about dynamic mode. maybe we need just
to set maximum limit and later create them dynamically with some step?
so if free threads is less then 100 we are creating next 1000 threads
or 2000 threads, if for last 10 minutes or 20 minutes - we have more
than 2000 threads free we are closing them - upto 1000 free. so tuning
will allow to change threads dynamically and performance, but in the
same time all time have them already pre created. I see you have some
stats - is current async threads value accessable in it? then it's
pretty simple will be to scan this value periodically and to
change async value dynamically with - external near running script.

soon I'am going to do benchmark with commet threads so when
40-50-80-100K threads long running. right now I was available to get
49500 opened threads with gevent and 45000 threads with uwsgi with the
same open file limit, not sure yet why so difference.

and for sure will try any new patches if you have another ideas how to
improve async performance.

P.S: also if processes=2 and async=1000 - it will mean 2000
threads created - right?

-=NGINX STATIC PAGE=-
Requests per second:    27423.47 [#/sec] (mean)

-=UWSGI 0.9.6.6=-
sync mode
Requests per second:    22635.61 [#/sec] (mean)

async 100
Requests per second:    21551.03 [#/sec] (mean)

async 40000
Requests per second:    6508.24 [#/sec] (mean)

async 80000
Requests per second:    1620.11 [#/sec] (mean)


-=UWSGI 0.9.6.6 better_async.patch applied=-
async 100
Requests per second:    21650.39 [#/sec] (mean)

async 40000
Requests per second:    6676.47 [#/sec] (mean)

async 80000
Requests per second:    1536.41 [#/sec] (mean)

P.S.: benches are much higher as they was done on dedicated server

And awesome work Roberto for sync mode - only 1/4 slower than static
nginx - looks very fast to me :)

_______________________________________________
uWSGI mailing list
[email protected]
http://lists.unbit.it/cgi-bin/mailman/listinfo/uwsgi

Reply via email to