curl-loader-devel

web server optimization question

[jari p. <jar...@ho...>]

First thank you for an excellent tool!

slightly OT question:

My main purpose for using  curl-load is to load a device (stateful fw)=20
beetween clients and server.

The question: how to optimize Apache and curl-load query parameters=20
to achieve  max in terms of CAPS when retrieving small files from one httpd=
  server.

Now with 1000 curl clients to Apache server I can  get max 1200-1300 CAPS.
in this test all clients are fetching the same 1k file. curl load config is=
 based on sample
10K example, but with 1000 clients.

  -how to optimize Apache (or would Lighthttpd be better ?)
  - what curl-load client options could be adjusted one increase connection=
 rate?
=20


(again for  this particular test I want to maximaze Connections/sec, and no=
t necessary
 throughput or concurrent connections )

Jari

_________________________________________________________________
Invite your mail contacts to join your friends list with Windows Live Space=
s. It's easy!
http://spaces.live.com/spacesapi.aspx?wx_action=3Dcreate&wx_url=3D/friends.=
aspx&mkt=3Den-us=

Re: web server optimization question

[Robert I. <cor...@gm...>]

Hi Jari,

On 9/2/07, jari pietila <jar...@ho...> wrote:
> My main purpose for using  curl-load is to load a device (stateful fw)
> beetween clients and server.
>
> The question: how to optimize Apache and curl-load query parameters
> to achieve  max in terms of CAPS

Default Apache configuration is really limiting.

>   -how to optimize Apache (or would Lighthttpd be better ?)

I would suggest you to use nginx, which was found even better,
than lighttpd. It has a Debian package and nice docs here, but you can use just
the default conf
http://wiki.codemongers.com/Main?action=show&redirect=Nginx

>   - what curl-load client options could be adjusted one increase connection
> rate?

We seen, that you do not need much connections to reach high CAPS.
Something between 100 and 1000 is enough.

Compile curl-loader with optimization as described:
http://curl-loader.sourceforge.net/doc/faq.html#big-load
http://curl-loader.sourceforge.net/high-load-hw/index.html
Most probably, you do not need to fine-tune linux, whereas you can try.

What is most important is to ensure persistance of the tcp-connections.
curl-loader makes its best, therefore make sure, that server does as well.

If your client machine has multiple-core or multiple CPU, you can also try
an option to run it with -t <number> option, where the numbers to try are
2, 4, 8

With nginx I have seen 6000 - 8000 CAPS with a small file of 100-200 bytes
and 200-400 curl-loader virtual clients, each with an established keep-alived
connection.

Highest number of CAPS you'll get as a guess, when
TIMER_URL_COMPLETION = 0
TIMER_AFTER_URL_SLEEP = 0

Best wishes and report to the list you best experience.

-- 
Sincerely,
Robert Iakobashvili,
coroberti %x40 gmail %x2e com
...........................................................
http://curl-loader.sourceforge.net
A web testing and traffic generation tool.

RE: web server optimization question

[jari p. <jar...@ho...>]

Robert,

I  did some experimenting and can share the results so far:
=20
- with ligthttpd and minimal server tuning, I got around 3300 CAPS with 100=
0 virtual clients.
    (load-curl  0.41 re-compiled with optimization) compared to 1200 with A=
pache.
  =20
      Then the good news:
      My client in this test was is dual-core P4 2.8Ghz, 1GB RAM, Fedora 5=
=20
     and with -t 2 option there was almost linear scaling leading about 600=
0 CAPS!
      Very impressive,  thanks  for  pointing  this  out.
   =20
     - I then switched to nginx, and  was able to reach  8000 CAPS  with -t=
 2 and 1000 virtual clients.
   (no change in client configuration )

  But there is also a change in server behavior, lighttpd does NOT use pers=
istent connections
   and closes each TCP connection (TCP FIN server ->client ).
   This is actually the kind of behavior, I wanted to see:
    NO persistence, thus  creating load on DUT that needs to track TCP sess=
ion creation and closing.
    =20
   Nginx  with default settings has persistence,
    (and not sure if there is directive to make it not to do so),
    so this seems to explain the higher CAPS number.

   I also experimented a little with the number of virtual clients,=20
   but I got about the same results with 256 clients,
    and performance started to degrade around 1500 clients.
   =20
       - Finally, I tried configuration  where client closing the connectio=
n, but with this=20
     I got  much lower CAPS with both lighttpd and nginx.

-jari=20




_________________________________________________________________
Connect to the next generation of MSN Messenger=A0
http://imagine-msn.com/messenger/launch80/default.aspx?locale=3Den-us&sourc=
e=3Dwlmailtagline=

RE: web server optimization question

[jari p. <jar...@ho...>]

Robert,

I  did some experimenting and can share the results so far:
=20
- with ligthttpd and minimal server tuning, I got around 3300 CAPS with 100=
0 virtual clients.
    (load-curl  0.41 re-compiled with optimization) compared to 1200 with A=
pache.
 =20
Then the good news:
 My client in this test was is dual-core P4 2.8Ghz, 1GB RAM, Fedora 5
 and with -t 2 option there was almost linear scaling leading about 6000 CA=
PS!
 Very impressive,  thanks  for  pointing  this  out.
  =20
  - I then switched to nginx, and  now got  8000 CAPS  with -t 2 and 1000 v=
irtual clients.
   (no change in client configuration )

But there was  also a change in server behavior, lighttpd does NOT use pers=
istent connections
and closes each TCP connection (TCP FIN server ->client ).
This is actually the kind of behavior, I wanted to see:
NO persistence, thus  creating load on DUT that needs to track TCP session =
creation and closing.
   =20
Nginx  with default settings has persistence,  (and not sure if there is di=
rective to make it not to do so),
so this seems to explain the higher CAPS number.

I also experimented a little with the number of virtual clients, but I got =
about the same results with 256 clients,
and performance started to degrade around 1500 clients.

- Finally, I tried configuration  where client closes the connection,=20
but with this  I got  much lower CAPS with both lighttpd and nginx.


_________________________________________________________________
Discover the new Windows Vista
http://search.msn.com/results.aspx?q=3Dwindows+vista&mkt=3Den-US&form=3DQBR=
E=

Re: web server optimization question

[Robert I. <cor...@gm...>]

Hi Jari,

On 9/7/07, jari pietila <jar...@ho...> wrote:

>  My client in this test was is dual-core P4 2.8Ghz, 1GB RAM, Fedora 5
>  and with -t 2 option there was almost linear scaling leading about 6000
> CAPS!
>   - I then switched to nginx, and  now got  8000 CAPS  with -t 2 and 1000
> virtual clients.
>

The numbers are matching our experience.

I also experimented a little with the number of virtual clients, but I got
> about the same results with 256 clients,
> and performance started to degrade around 1500 clients.
>

Y, indeed to get more CAPS , its about the same with 200 - 1000 clients

- Finally, I tried configuration  where client closes the connection,
> but with this  I got  much lower CAPS with both lighttpd and nginx.
>

It is also matching our experiments.

If you see idle CPU remaning at your loading machine, you can try adding
more servers. Else, you can scale either by increasing the CPUs/memory
of the client and server machines or by increasing their numbers.
http://curl-loader.sourceforge.net/high-load-hw/index.html

Thanks and best wishes.

Sincerely,
Robert Iakobashvili,
coroberti %x40 gmail %x2e com
...........................................................
http://curl-loader.sourceforge.net
A web testing and traffic generation tool.

Re: web server optimization question

[Aleksandar L. <al-...@no...>]

Hi,

On Fre 07.09.2007 03:12, jari pietila wrote:
>Robert,
>
>I  did some experimenting and can share the results so far:

Nice to share your experience ;-), thanks.

>- with ligthttpd and minimal server tuning, I got around 3300 CAPS with
>   1000 virtual clients.
>    (load-curl  0.41 re-compiled with optimization) compared to 1200
>    with Apache.
>   
>      Then the good news:
>      My client in this test was is dual-core P4 2.8Ghz, 1GB RAM, Fedora 5 
>     and with -t 2 option there was almost linear scaling leading about 6000 CAPS!
>      Very impressive,  thanks  for  pointing  this  out.
>    
>     - I then switched to nginx, and  was able to reach  8000 CAPS
>       with -t 2 and 1000 virtual clients.
>       (no change in client configuration )

Cool ;-)

>  But there is also a change in server behavior, lighttpd does NOT use
>  persistent connections and closes each TCP connection (TCP FIN server
>  ->client ).
>   This is actually the kind of behavior, I wanted to see:
>    NO persistence, thus  creating load on DUT that needs to track TCP
>    session creation and closing.

I'am not a fan of lighty but to be fair have you looked into:

http://trac.lighttpd.net/trac/wiki/Docs%3APerformance

>   Nginx  with default settings has persistence,
>    (and not sure if there is directive to make it not to do so),
>    so this seems to explain the higher CAPS number.

Afaik there is no option to switch off the keep-alive but you can tune
it:

http://wiki.codemongers.com/NginxHttpCoreModule#keepalive_timeout

>   I also experimented a little with the number of virtual clients, 
>   but I got about the same results with 256 clients,
>    and performance started to degrade around 1500 clients.
>    
>       - Finally, I tried configuration  where client closing the connection, but with this 
>     I got  much lower CAPS with both lighttpd and nginx.

Have you tried to turn on/off tcp_{nodelay,nopush), yust for info?!

http://wiki.codemongers.com/NginxHttpCoreModule#tcp_nodelay

BR

Aleks

Re: web server optimization question

[Robert I. <cor...@gm...>]

Aleks,

On 9/7/07, Aleksandar Lazic <al-...@no...> wrote:
>
> I'am not a fan of lighty but to be fair have you looked into:
>
> http://trac.lighttpd.net/trac/wiki/Docs%3APerformance
>
> >   Nginx  with default settings has persistence,
> >    (and not sure if there is directive to make it not to do so),
> >    so this seems to explain the higher CAPS number.
>
> Afaik there is no option to switch off the keep-alive but you can tune
> it:
>
> http://wiki.codemongers.com/NginxHttpCoreModule#keepalive_timeout
>
> >   I also experimented a little with the number of virtual clients,
> >   but I got about the same results with 256 clients,
> >    and performance started to degrade around 1500 clients.
> >
> >       - Finally, I tried configuration  where client closing the
> connection, but with this
> >     I got  much lower CAPS with both lighttpd and nginx.
>
> Have you tried to turn on/off tcp_{nodelay,nopush), yust for info?!
>
> http://wiki.codemongers.com/NginxHttpCoreModule#tcp_nodelay
>

Interesting links, thanks.

-- 
Sincerely,
Robert Iakobashvili,
coroberti %x40 gmail %x2e com
...........................................................
http://curl-loader.sourceforge.net
A web testing and traffic generation tool.