NXWEB - ultra-fast web server based on libev

Yaroslav yarosla at gmail.com
Sun Dec 4 01:25:27 CET 2011 

Hi, everybody!

I tried to use weighttp for benchmarking but encountered some problems when
using large number of concurrent connections in multi-threaded mode. For
example:

   weighttp -c 10000 -n 100000 -t 4 -k http://...

This almost always produces quite a number of '104 Connection reset by
peer' errors, many connections do not close, which means the test runs
long, and measurements get poor. This is not specific to nxweb, same
situation with all tested servers. I even tried it on different hardware
with CentOS. Same situation. The interesting part is that when used in
single thread mode (-t 1) there are no such errors even with large number
of connections.

I've spent whole day digging the source of weighttp, even made some
patches, but could not figure out the source of the problem.

Finally I've decided to write my own benchmarking tool, and successfully
did it. With the same result though... As I also used libev.

Anyway I made some improvements over weighttp, namely:
- load is evenly distributed among threads, in weighttp one thread could
finish much earlier than others
- I forcibly cut timed out connections, not allowing them to spoil the test

Otherwise it is the same efficient as weighttp. Need to mention that I have
not implemented (yet) chunked encoding, custom headers and IPv6, which
weighttp implements.

Here is the project page if anyone interested:
https://bitbucket.org/yarosla/httpress/

I plan to redo all benchmarks for NXWEB in a few days. With the new tool
results are much better now.

I wonder if anybody experienced same issues with hanging/reset connections
with libev under high concurrency?

Regards,

Yaroslav

On Fri, Dec 2, 2011 at 3:14 PM, Yaroslav <yarosla at gmail.com> wrote:

> Hi Mickael,
>
> I mentioned in limitations that it has only been tested on Linux. I can't
> test it on OSX unfortunately, and frankly speaking portability was not my
> primary goal. At least not at this stage. Hosting providers mostly offer
> Linux, never heard about OSX.
>
> But still I think it should be possible to port it to OSX, it's not
> Windows anyway.
>
> What causing compile errors is the absense of argp and obstack in OSX's
> library. argp is not so critical for NXWEB - this is just command line
> parsing. While obstack is used extensively by the core.
>
> Thanks for pointing to benchmarking tool. I will take a look at it. I
> wasn't satisfied by ab, but couldn't find anything better. httperf happened
> to be too complicated for my brain.
>
> Yaroslav
>
>
> On Fri, Dec 2, 2011 at 2:56 PM, Zabrane Mickael <zabrane3 at gmail.com>wrote:
>
>> Hi Yaroslav,
>>
>> My first try to compile NxWeb (OSX 10.6.8 + libev4):
>>
>> *nxweb* $  make
>> mkdir -p bin/Release
>> mkdir -p obj/Release/nxweb
>> gcc -c -o obj/Release/nxweb/main.o nxweb/main.c -pthread
>> -Wno-strict-aliasing -O2 -s
>> nxweb/main.c:29:18: error: argp.h: No such file or directory
>> In file included from nxweb/nxweb_internal.h:30,
>>                  from nxweb/main.c:35:
>> nxweb/nxweb.h:30:21: error: obstack.h: No such file or directory
>> In file included from nxweb/nxweb_internal.h:30,
>>                  from nxweb/main.c:35:
>> nxweb/nxweb.h:121: error: field ‘data’ has incomplete type
>> nxweb/nxweb.h:122: error: field ‘user_data’ has incomplete type
>> nxweb/main.c:123: error: array type has incomplete element type
>> nxweb/main.c:140: error: expected ‘=’, ‘,’, ‘;’, ‘asm’ or ‘__attribute__’
>> before ‘parse_opt’
>> nxweb/main.c:165: error: variable ‘argp’ has initializer but incomplete
>> type
>> nxweb/main.c:165: warning: excess elements in struct initializer
>> nxweb/main.c:165: warning: (near initialization for ‘argp’)
>> nxweb/main.c:165: error: ‘parse_opt’ undeclared here (not in a function)
>> nxweb/main.c:165: warning: excess elements in struct initializer
>> nxweb/main.c:165: warning: (near initialization for ‘argp’)
>> nxweb/main.c:165: warning: excess elements in struct initializer
>> nxweb/main.c:165: warning: (near initialization for ‘argp’)
>> nxweb/main.c:165: warning: excess elements in struct initializer
>> nxweb/main.c:165: warning: (near initialization for ‘argp’)
>> make: *** [obj/Release/nxweb/main.o] Error 1
>>
>>
>> I think you should update your *benchmark* to use *weighttp*:
>> http://redmine.lighttpd.net/projects/weighttp/wiki
>> As pointed by G-WAN's author, *ab* performs badly on multi-cores.
>>
>> Regards,
>> Zabrane
>>
>> On Dec 2, 2011, at 9:58 AM, Yaroslav wrote:
>>
>> What a shame. Sorry about that. It is my first public repo at Bitbucket,
>> I turned 'private' flag off, but did not check that I have to make wiki and
>> issue tracker public as well.
>>
>> It should be all OK now.
>>
>> Sorry for the inconvenience.
>>
>> On Fri, Dec 2, 2011 at 12:13 PM, Chris Brody <chris.brody at gmail.com>wrote:
>>
>>> Yeah I created an account with Bitbucket and still could not get access
>>> to the Wiki.
>>>
>>> Can you please make this project public if you want to announce it?
>>>
>>>
>>> On Fri, Dec 2, 2011 at 3:04 AM, Rodrigo Campos <rodrigo at sdfg.com.ar>wrote:
>>>
>>>> On Fri, Dec 02, 2011 at 03:55:27AM +0400, Yaroslav wrote:
>>>> > Hi,
>>>> >
>>>> > I'd like to thank the authors of libev for very good product. Based
>>>> on it I
>>>> > have developed NXWEB web server, which I'd like to present to
>>>> community. Here
>>>> > are the highlights:
>>>> >
>>>> > NXWEB project page: https://bitbucket.org/yarosla/nxweb/
>>>>
>>>> Sounds interesting, but the project page asks for a login. It seems I
>>>> can not
>>>> see anything without an account :(
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> Thanks,
>>>> Rodrigo
>>>>
>>>> _______________________________________________
>>>> libev mailing list
>>>> libev at lists.schmorp.de
>>>> http://lists.schmorp.de/cgi-bin/mailman/listinfo/libev
>>>
>>>
>>>
>>> _______________________________________________
>>> libev mailing list
>>> libev at lists.schmorp.de
>>> http://lists.schmorp.de/cgi-bin/mailman/listinfo/libev
>>>
>>
>> _______________________________________________
>> libev mailing list
>> libev at lists.schmorp.de
>> http://lists.schmorp.de/cgi-bin/mailman/listinfo/libev
>>
>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.schmorp.de/pipermail/libev/attachments/20111204/cc200ec5/attachment.html>

More information about the libev mailing list