Can't remember. I really think that the people with interesting perfs data, if they don't consider their data as confidential, should share that on the wiki. Everybody around is asking for scalability data, and there is no real source of information. David Ponzone Direction Technique email: david.ponzone@ipeva.fr tel: 01 74 03 18 97 gsm: 06 66 98 76 34 Service Client IPeva tel: 0811 46 26 26 www.ipeva.fr - www.ipeva-studio.com Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur. Le 20/08/2010 à 12:55, Jean-Yves F. Barbier a écrit :

Le Fri, 20 Aug 2010 12:36:46 +0200, David Ponzone <david.ponzone@ipeva.fr> a écrit :

...

...

But some people told me recently on #freeswitch that they had achieved close to 300cps on only one SIP profile.

Hi Dave, do you remember on what kind of CPU (& RAM qty) it was?

JY -- How to Raise Your I.Q. by Eating Gifted Children -- Book title by Lewis B. Frumkes

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

I think it should be doable as a table. I created a template table here: http://wiki.freeswitch.org/wiki/Performance_testing_and_configurations#Some_... If you think something is missing, please feel free to tell me or update the table yourself. David Ponzone Direction Technique email: david.ponzone@ipeva.fr tel: 01 74 03 18 97 gsm: 06 66 98 76 34 Service Client IPeva tel: 0811 46 26 26 www.ipeva.fr - www.ipeva-studio.com Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur. Le 20/08/2010 à 13:47, Jean-Yves F. Barbier a écrit :

Le Fri, 20 Aug 2010 13:22:24 +0200, David Ponzone <david.ponzone@ipeva.fr> a écrit :

...

Can't remember.

I really think that the people with interesting perfs data, if they don't consider their data as confidential, should share that on the wiki. Everybody around is asking for scalability data, and there is no real source of information.

Hey good idea! - as far as modus operandi is fully exposed.

-- Jesus Saves, Moses Invests, But only Buddha pays Dividends.

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

JY, There is a misunderstanding. I only added the table. The DB part was written before by someone else. Check the history of the page to see who wrote that, or correct it yourself. David Ponzone Direction Technique email: david.ponzone@ipeva.fr tel: 01 74 03 18 97 gsm: 06 66 98 76 34 Service Client IPeva tel: 0811 46 26 26 www.ipeva.fr - www.ipeva-studio.com Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur. Le 20/08/2010 à 16:34, Jean-Yves F. Barbier a écrit :

Le Fri, 20 Aug 2010 15:37:22 +0200, David Ponzone <david.ponzone@ipeva.fr> a écrit :

...

I think it should be doable as a table. I created a template table here: http://wiki.freeswitch.org/wiki/Performance_testing_and_configurations#Some_...

Hehe not so bad after all for such a venerable CPU :) May be the sipp command line used would help to standardize tests.

One thing about the DB part though: you should also mention using another file system for the DB's files on HD or even SSD, AND warn people about total data loss in case of power failure w/ the tmpfs possibility.

JY -- Iowa State -- the high school after high school! -- Crow T. Robot

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

The problem is that performance testing is very fluid and relevant to each machine and OS and test performed and you can't get concrete expectations without some advanced skills in deploying FS. On Fri, Aug 20, 2010 at 9:53 AM, David Swardstrom <dswardstrom@remotelink.com> wrote:

I agree that it would be a good idea to share performance data on the Wiki. If this is done, I have some suggestions (after trying to walk the Wiki many times): - Create a "main" page for performance data. - Allow people to build their own pages relating to performance. - Link these pages to this main page. - Link the main performance page on the Main Page of the Wiki (http://wiki.freeswitch.org/wiki/Main_Page).  Perhaps as a new bullet item in: Feature Documentation

Then find any already provided performance data in the Wiki pages and get them linked in.

We (RemoteLink) have some data regarding number of Conference Users that we could share.

-- View this message in context: http://freeswitch-users.2379917.n2.nabble.com/freeswitch-CPU-usage-tp5443765... Sent from the freeswitch-users mailing list archive at Nabble.com.

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

-- Anthony Minessale II FreeSWITCH http://www.freeswitch.org/ ClueCon http://www.cluecon.com/ Twitter: http://twitter.com/FreeSWITCH_wire AIM: anthm MSN:anthony_minessale@hotmail.com GTALK/JABBER/PAYPAL:anthony.minessale@gmail.com IRC: irc.freenode.net #freeswitch FreeSWITCH Developer Conference sip:888@conference.freeswitch.org googletalk:conf+888@conference.freeswitch.org pstn:+19193869900

Le Fri, 20 Aug 2010 10:10:32 -0500, Anthony Minessale <anthony.minessale@gmail.com> a écrit :

The problem is that performance testing is very fluid and relevant to each machine and OS and test performed and you can't get concrete expectations without some advanced skills in deploying FS.

Well, it could have a real interest by adding the tweaked parameters, may be there are some, yet unexplored, that would raise the response. -- Why don't elephants eat penguins ? Because they can't get the wrappers off ...

Le Fri, 20 Aug 2010 11:01:18 -0500, Anthony Minessale <anthony.minessale@gmail.com> a écrit :

Its an ok thing as long as it comes with a disclaimer that "results may vary" and "you have to know what you are doing" And those who work on it need to volunteer to help with the emails it may generate asking why they can't match the results =D

It can be resumed: YMMV, RTFM & FYI... --

I already removed the stress tests part from the table. I will try to add other information that was given here. For the disclaimer part, as english is not my native language, I would appreciate if someone could recommend one with the right content, thus avoiding our dear FreeSWITCH devs to go to jail in the future :) David Ponzone Direction Technique email: david.ponzone@ipeva.fr tel: 01 74 03 18 97 gsm: 06 66 98 76 34 Service Client IPeva tel: 0811 46 26 26 www.ipeva.fr - www.ipeva-studio.com Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur. Le 20/08/2010 à 18:21, François Delawarde a écrit :

Stress test results are only useful to make the headlines or for PhD students, those tests should only be made by experts (unbiased if possible) that and results should be shown in a special "cool-pure-raw performance testing" page.

For the rest of us, real life scenarios with a small description of the modules involved would be useful to help with server dimensioning. If people contribute, with enough data it could give a pretty nice idea of what to expect.

Of course those results can vary a lot so a huge disclaimer must be there to avoid having special people (mainly US citizens) suing the wiki owner. ;-)

François.

On Fri, 2010-08-20 at 11:01 -0500, Anthony Minessale wrote:

...

Its an ok thing as long as it comes with a disclaimer that "results may vary" and "you have to know what you are doing" And those who work on it need to volunteer to help with the emails it may generate asking why they can't match the results =D

On Fri, Aug 20, 2010 at 10:52 AM, Jean-Yves F. Barbier <12ukwn@gmail.com

...

wrote: Le Fri, 20 Aug 2010 10:10:32 -0500, Anthony Minessale <anthony.minessale@gmail.com> a écrit :

...

The problem is that performance testing is very fluid and relevant to each machine and OS and test performed and you can't get concrete expectations without some advanced skills in deploying FS.

Well, it could have a real interest by adding the tweaked parameters, may be there are some, yet unexplored, that would raise the response.

-- Why don't elephants eat penguins ? Because they can't get the wrappers off ...

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

These are all set correctly. I just tried some older version ( 1.0.4 ) and found that I could run up to 600 cps. I am very surprised that the performance seems to be degraded with the newer version. On Fri, Aug 20, 2010 at 6:36 PM, David Ponzone <david.ponzone@ipeva.fr> wrote:

Have you followed all the recommendations for best performances ? Sorry if that sounds obvious to you but things like: ulimit -s 240 reducing core loglevel to warning increasing tx/rx buffer of the ethernet to 4096 Are you hitting only one SIP profie with your stress test ? Remember there is only one thread per SIP profile. But some people told me recently on #freeswitch that they had achieved close to 300cps on only one SIP profile. David Ponzone  Direction Technique email: david.ponzone@ipeva.fr tel:      01 74 03 18 97 gsm:   06 66 98 76 34 Service Client IPeva tel:      0811 46 26 26 www.ipeva.fr  -   www.ipeva-studio.com Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur.

Le 20/08/2010 à 11:29, Woody Dickson a écrit :

Hi,

I am doing some experiments with Freeswitch by torturing it to see how the machine's CPU response to heavy loaded situation. The test is done on a 16 core 5550 dual quad core server running fedora 2.6.30.10-105.2.23.fc11.x86_64 OS.

What I found so strange was that while CPU usage remains pretty low and distributed among all cores at 190 - 200 calls per second.  Then, after added a few more calls per second, all CPU becomes fully utilized.

Is this due to some wrong setting?  Any idea how I can tweak the configuration and continue my test?

Thanks, Woody

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

Le Fri, 20 Aug 2010 09:52:21 -0500, Anthony Minessale <anthony.minessale@gmail.com> a écrit :

Yet another load test thread.........................

This is why we don't like them.

Yep, but people are like that: we looove sooo much tests! (any volunteer to write mod_test for FS and mod_test_test to test mod_test?:) ...

if you insist on sipp use our scenario file that fixes several errors in the default one.

http://www.freeswitch.org/eg/load_test/dft_cap.xml

Hum, reminds me about quite the same values I saw in an OpenSER serie of tests, thanks for a clear standardization Anthony. JY -- Trust no one -- The X-Files

On 20/8/2010 5:52 μμ, Anthony Minessale wrote:

Yet another load test thread.........................

This is why we don't like them.

Too bad you missed my presentation at cluecon.....

Do you hear yourself btw? Quibbling about completely free software *only* doing 200cps in your fake test? I don't know if I missed any message in this thread, but why do you say that the tester did a fake test?

200cps is 4 times the industry standard FYI and if you were really doing 200cps in real life you would be so rich at our expense that you could afford more boxes.

you clearly ignored all of our repeated recommendations. Use the following or you are on your own.......

This OS Centos 5.x x86_64

If someone is on his own for any other OS than Centos 5.x x86_64 then why is FS offered for so many different platforms? It's only natural for people to want to test it on their favorite platform... Perhaps, as you recommend, Centos 5.x x86_64 gives the best results over all other platforms, but last time I checked, a user is still allowed to ask performance related questions for his favorite platform in this mailing list, right? Woody's post was not expressing any negative opinion about FS's performance. He didn't comment on whether 200cps was too little or too much. He just wrote that the CPU usage increased in an unexpected manner and wonders how/if he can solve this. He wrote that he was doing a load test scenario, he didn't write that he was making money at 200cps "at your expense". And he wrote all this in a manner which was neither insulting nor abusive. The poster just wanted to learn something, see if others have gotten better results with other setups. He was trying to make a valid discussion to solve a problem, and other users were actually replying to him. This is what mailing lists are for, right? I really don't understand why red flags have to be raised every time the word "performance" is mentioned in this mailing list. -- Best regards, Vlasis Hatzistavrou.

That's easy for you to say because you are using the mailing list not running it or taking a personal vested interest in maintaining the quality of it's content. We are trying to make sure we help everyone all day long every day constantly. If you personally had to deal with every time we get this recurring "load testing" threads, you might understand. The alternative would be for us to lower our standards and let the list run away with itself and not care if we actually keep up on it. What mailing lists are for is anything the owner of the list wants it to be for. He is welcome to continue to say anything he wants and we are obliged to reply however we want. It's a 2 way street. You seem to suggest that we don't have to right to have policies on our own community that many people are very satisfied with only because we do exactly what we want to run our own community. You as well are welcome to your opinion which is duly noted. We still however, do not like load testing threads. The individuals here who have volunteered to set up some wiki info with the proper disclaimers may produce a resource we can send people to and be done with it. That is what WIKIs and FAQs are for..... On Fri, Aug 20, 2010 at 11:19 AM, Vlasis Hatzistavrou <vhatz@kinetix.gr> wrote:

 On 20/8/2010 5:52 μμ, Anthony Minessale wrote:

...

Yet another load test thread.........................

This is why we don't like them.

Too bad you missed my presentation at cluecon.....

Do you hear yourself btw? Quibbling about completely free software *only* doing 200cps in your fake test? I don't know if I missed any message in this thread, but why do you say that the tester did a fake test?

...

200cps is 4 times the industry standard FYI and if you were really doing 200cps in real life you would be so rich at our expense that you could afford more boxes.

you clearly ignored all of our repeated recommendations. Use the following or you are on your own.......

This OS Centos 5.x x86_64

If someone is on his own for any other OS than Centos 5.x x86_64 then why is FS offered for so many different platforms?

It's only natural for people to want to test it on their favorite platform... Perhaps, as you recommend, Centos 5.x x86_64 gives the best results over all other platforms, but last time I checked, a user is still allowed to ask performance related questions for his favorite platform in this mailing list, right?

Woody's post was not expressing any negative opinion about FS's performance. He didn't comment on whether 200cps was too little or too much. He just wrote that the CPU usage increased in an unexpected manner and wonders how/if he can solve this. He wrote that he was doing a load test scenario, he didn't write that he was making money at 200cps "at your expense". And he wrote all this in a manner which was neither insulting nor abusive.

The poster just wanted to learn something, see if others have gotten better results with other setups. He was trying to make a valid discussion to solve a problem, and other users were actually replying to him.

This is what mailing lists are for, right?

I really don't understand why red flags have to be raised every time the word "performance" is mentioned in this mailing list.

-- Best regards, Vlasis Hatzistavrou.

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

-- Anthony Minessale II FreeSWITCH http://www.freeswitch.org/ ClueCon http://www.cluecon.com/ Twitter: http://twitter.com/FreeSWITCH_wire AIM: anthm MSN:anthony_minessale@hotmail.com GTALK/JABBER/PAYPAL:anthony.minessale@gmail.com IRC: irc.freenode.net #freeswitch FreeSWITCH Developer Conference sip:888@conference.freeswitch.org googletalk:conf+888@conference.freeswitch.org pstn:+19193869900

I am not accusing the user of anything I am making factual statements. They are indeed fake tests because it's sipp and 600cps. I said *if* he was doing many hundreds of cps *in real life* he would indeed be very rich at our expense, (meaning using our software not meaning he is chaining us to a wall in his basement) and thus, he would be able to afford more boxes. Do you feel I was not willing to help when I provided a slew of configuration options that will increase performance as well as proper sipp files to use? I also endorsed the idea of the wiki page as long as they maintain it. Why are you arguing on someone else's behalf on a matter that needs no arguing. You are continuing to argue that I am not allowed to say what I want on my own mailing list which is somewhat troubling. If you decide to reply again, please limit the length of your post. On Fri, Aug 20, 2010 at 5:17 PM, Vlasis Hatzistavrou (KTI) <vhatz@kinetix.gr> wrote:

Anthony Minessale wrote:

...

That's easy for you to say because you are using the mailing list not running it or taking a personal vested interest in maintaining  the quality of it's content.

We are trying to make sure we help everyone all day long every day constantly. If you personally had to deal with every time we get this recurring "load testing" threads, you might understand.

I understand that you may feel frustrated on a topic that you have already covered many times in the past in the list.

The fact, however, that the list still receives performance-related questions shows that the users are still interested in the subject, that they still have questions and that they need to compare practices, information, set ups, anything one can think of, despite the information already available in the wiki and in the mailing list.

The fact that some users still ask and wonder about performance does not mean that they dispute FS's performance or that the users are too lazy to test thoroughly or that they don't know what they are doing (OK, granted some of us don't, but may be willing to learn).

Performance related questions pop up all the time in the mailing lists of probably every software that I have watched. IMHO, this is a good thing. Higher performance is of interest in all kinds of software, this being especially true for VoIP software.

The question

"ok, great, we get 200 cps. Now, how can we get the software to do more"

does not show stupidity, greed or ingratitude. It is a valid question that can move forward the development of any software.

...

The alternative would be for us to lower our standards and let the list run away with itself and not care if we actually keep up on it.

What mailing lists are for is anything the owner of the list wants it to be for. He is welcome to continue to say anything he wants and we are obliged to reply however we want. It's a 2 way street.  You seem to suggest that we don't have to right to have policies on our own community that many people are very satisfied with only because we do exactly what we want to run our own community.

I never suggested that you don't have the right to have policies in the list that you have created. I'm saying something different: although I didn't see anywhere a policy that the list users are forbidden to post performance related questions, this is exactly how such questions are treated in practice.

Of course, I can't even begin to imagine any software mailing list with a policy of forbidding the discussion of "taboo" subjects. And this is exactly the reason why I wrote my post.

Specifically in this case, Woody wrote a valid question about a subject not forbidden by any policy. He deserved an answer by anyone willing to reply as much as the next user on the list.

I know that you have a busy schedule, and that you cannot possibly participate in all threads, especially in subjects already covered, even more so for subjects that you don't like.

But there are already experienced users in the list who are willing to reply. Some users, who were willing to help, actually did reply and in fact a whole discussion about performance and test data was generated, showing beyond doubt that this is a hot topic for FS users. And IMHO this is a good thing for FS.

The original question in this thread did not attack your work or FreeSWITCH in general in any way. So, no need to accuse the user of making money on your expense or that he performs fake tests. It was just a question, after all.

Best regards, Vlasis Hatzistavrou.

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

-- Anthony Minessale II FreeSWITCH http://www.freeswitch.org/ ClueCon http://www.cluecon.com/ Twitter: http://twitter.com/FreeSWITCH_wire AIM: anthm MSN:anthony_minessale@hotmail.com GTALK/JABBER/PAYPAL:anthony.minessale@gmail.com IRC: irc.freenode.net #freeswitch FreeSWITCH Developer Conference sip:888@conference.freeswitch.org googletalk:conf+888@conference.freeswitch.org pstn:+19193869900

<snip>

You wrote in your reply to Woody: "Do you hear yourself btw?" I can ask you the same question: Do you hear yourself? You act as if some subjects are taboo in this list.

Taboos aren't really the issue. Spending valuable time and energy in unfruitful debates is an issue.

...

If you decide to reply again, please limit the length of your post.

Now it is the size of my posts that troubles you? I've seen longer posts in the list and you didn't complain.

Brevity is the essence of communication - I highly recommend it.

Perhaps you should add a size restriction directive in the mailing list policy, along with a "no performance questions allowed" directive. That should clear things up significantly.

Would you be willing just to let the subject go? I think everyone is tired of this thread. -MC P.S. - Just a clarification: Using SIPp is a "fake" test to do 200/400/600/xxx calls per second. Getting thousands of actual humans making actual calls is a "real" test. In other words, since SIPp can only do so much its test results are of limited value in the real world.

Vlasis, please calm down and try to understand the context you're interacting with. This is not a matter of civil rights or politically correctness or different sensibilities or respect for the freedom of speech or whatever thing of this kind. This is a matter about a consistent, practical and every time asserted displease from the core team developers (and most of the not-so-core team developers) toward stress tests and performance measurements. It has been stated many many times, and the reason is: because fake test (sipp) does not gives you results that reflects in real world case, because sipp is black magic, because there are many variable involved (OS, hardware, kind of load, etc), because this is likely to attract undue interest on naked figures [we can do 1 million cps!!!!!], because to get it right involves a tremendous lot of experience, because the only useful test for your usage is to put in production a reasonable dimensioned machine and then see of it goes, because we read all the message in the list and particularly on performance themes we are pushed to intervene before too many wrong things are asserted, we don't like threads debating about performance testing. But, because we're human and nice, and helpful and caring, on the side of the bashing we also offer pearls of wisdom that can be used for the purpose, instead of simply bashing or ignore the poster. So, take it as a mysterious ways of the Zen Masters to visit black art on performance test posters. -giovanni On Sat, Aug 21, 2010 at 2:41 PM, Vlasis Hatzistavrou (KTI) <vhatz@kinetix.gr> wrote:

Michael Collins wrote:

<snip>

...

You wrote in your reply to Woody: "Do you hear yourself btw?" I can ask you the same question: Do you hear yourself? You act as if some subjects are taboo in this list.

Taboos aren't really the issue. Spending valuable time and energy in unfruitful debates is an issue.

People at times have been accused of various different things in this list just because they dared to ask something! And all you have to say is that this is an unfruitful debate?

For me the whole debate is VERY fruitful if the next user who dares to mention the dreaded word "performance" is spared the irony and the bashing.

...

...

If you decide to reply again, please limit the length of your post.

Now it is the size of my posts that troubles you? I've seen longer posts in the list and you didn't complain.

Brevity is the essence of communication - I highly recommend it.

Then, please recommend it to all cases and not just in my case.

Tony's reply was not so brief either, so, please go ahead and say the same to him. That is, only if you care to appear impartial.

Me on the other hand, I'd prefer an actual reply to my arguments, instead of dismissing the entire discussion for the sake of brevity. Up to now I have seen more evasion tactics than arguments against the points I've made.

-MC

P.S. - Just a clarification: Using SIPp is a "fake" test to do 200/400/600/xxx calls per second. Getting thousands of actual humans making actual calls is a "real" test. In other words, since SIPp can only do so much its test results are of limited value in the real world.

Really, is this what a "fake" test is? Because I thought that it was a test that never actually happened. English is not my native language, but at least I know that much.

What you wrote in your P.S. is so wrong in so many ways. For the "sake of brevity" I will only say that it is a bad attempt to cover Tony's rude reply to a list user and I'll leave it at that for the moment.

Otherwise we will have to get in a discussion of what a test is, what is a "real" as opposed to a "fake" test and whether sipp or test tools in general have limited value in the real world. I don't have a problem with long discussions, but is seems they are not favored in this context so I'll pass for the moment, until the circumstances require it.

...

Perhaps you should add a size restriction directive in the mailing list policy, along with a "no performance questions allowed" directive. That should clear things up significantly.

Would you be willing just to let the subject go? I think everyone is tired of this thread.

Tired? Now you can speak for "everyone"? I know quite a few people who are actually reading this thread with lots of interest.

Sure, I'm willing to let go... Are you?

Have a great weekend, Vlasis.

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

-- Sincerely, Giovanni Maruzzelli Cell : +39-347-2665618

Le Sat, 21 Aug 2010 15:41:13 +0300, "Vlasis Hatzistavrou (KTI)" <vhatz@kinetix.gr> a écrit : PLS Vlasis, calm down, I think eveyone is loosing temper riht now. Let's say you took a test option, whether it is a good one or not could be debated for a long time, but if YOU think it is the right way to go for YOUR needs then go with it! So my advice would be to eventually adjust your method only if you think it would help in your particular case of operation; and stick to it from now in order to always use the same test environment, to have a constant relevance of results and thus constant comparisons possibilities too in time :) (nooo, not in the head!!!:) JY --

Le Sat, 21 Aug 2010 17:58:27 +0300, "Vlasis Hatzistavrou (KTI)" <vhatz@kinetix.gr> a écrit :

Hello Jean-Yves,

Thanks for the effort to calm me down, but I am calm and fine. :)

That I understand, Vlasis; just a matter of speaking ;)

Just for the record I wasn't the one doing the tests, I was just defending the guy who asked the original question because I've been in his place and felt for him.

My purpose was just to extinguish this sterile discussion. Each part have real good arguments to counter the other part, but as nobody wants to leave its part of the cake I'm trying to guess a solution that would be the best for everybody tout le monde. This is why I think even a "bad" test could be good BUT ONLY if test conditions are even every time it is performed. The *real* question behind all of this is: why always test and compare? Wouldn't it be because large and greedy companies (also ear governments, especially corrupted ones by these same large companies) put that in our mind to sell every day more (and "better") to people that don't need it... Think about that. When I'm working for other I try to make it for the best and do it like if it was for myself - so I don't ask less from open-source devs that make tremendous programs, which if fortunately (almost) the case. So I can make a proposition to the dev team to _*definitely*_ fire down this kinda non-discussion: * Make a realistic serie of recursive test (who's more skilled than you to build that?:), * Always use the same machine with the same environment (possibly a "weak" one, such as a simple P4 2.0GHz w/ 2GB RAM: think about third world countries for which this kind of project is very important), * Publish the test serie results each time it is needed. (nooo, not the head AND the kidneys!!!) JY PS: I didn't choosed the fortune signature, it is a cron:) -- There appears to be irrefutable evidence that the mere fact of overcrowding induces violence. -- Harvey Wheeler

Nyamul Hassan wrote:

One thing that you are totaly "ignoring" Vlasis, is that the performance tests are irritating to the core-devs because FreeSwitch's CPS is already so high, that their effort / time are better used at improving / adding other things. Like the new "concurrent call recovery" feature, which is just simply crazy, if you ask me.

Hassan, (no irony intended, please believe me) but one thing you are totally ignoring too is that we were not debating about the validity of the tests. We were talking about attitude towards posts with questions on certain subjects. If users are still talking about performance this is because they are interested in it and are drawn to FS because of it. It's only natural to talk about it. And users of all software are talking about performance.

Just think about it, at 200 CPS, for 100% ASR (IVR) and 1 min ALOC / ACD, you're talking about 12,000 concurrent sessions. Or in a "real-life" call routing scenario, at 50% ASR and 5 mins ALOC, it is equivalent to 30,000 concurrent sessions, or 1,000 E1s!

And all that, using regular, run-of-the-mill, processor, motherboard, etc. The server Woody mentions would probably cost around $ 6,000 - 7,000/- from HP or IBM. Who can beat that? And, to top that off, Woody was comparing to 600 CPS that he got on v1.04 (36,000 or 90,000 conc. sess.)!

The core-devs are very passionate about their work. And, they have created a beast! So, get the beast to work "in real life", and then complain about problems.

Hassan, you didn't understand Woody's post neither mine. Did you read me or Woody complaining about problems with FS? I don't think so. Woody asked a very simple question. He didn't complain. Yet he was bashed. There was no complain in his message to the list, unless he sent another one which I missed completely. On the other hand, I have 2 questions regarding your comments: 1) Where exactly did you read 600 cps? Perhaps I missed a post but I only read 1 post from Woody with a scenario of 190 to 200 cps. Still a very high score, but this is not the point. 2) Sorry if it sounds ironic (again, no irony intended) but why are do you suggest that we complain about problems only after we use FS in real life? Because many of us do. On the other hand, why are lab tests that much frowned upon? Some people know how to test and reproduce production scenarios in a lab. Really.

I've seen them instantly jump into helping with the situation. Bug reports are fixed within hours. Even the "latest GIT" is claimed to be the most stable versions! Not many software projects (open source or not) can boast that, if at all. Hassan, no-one questioned the hard work put in by the developers. We were having a different discussion here.

As a sidenote, when the latest git/svn which, by nature, is work under development, is more stable than the latest stable release then yes, not many projects can boast that. On the contrary, most projects avoid this situation, because the users will have to put in production source code which is in a "fluid" state. Again, this is just a sidenote, (a comment on a comment) no intention to start a new flame here. :) Rgds, Vlasis.

Vlasis, Woody's 2nd email talked about 600 CPS on v1.04. And, perhaps you can share your usage scenario, and how much pressure you could put on FreeSwitch before it "crashed"? No hard feelings, and request to you is to also not take any from Tony's comments. The core team, as I've seen, is devoted to make FS better for "production use". Point out any flaw in "production", and you'll see it first hand. Will like to see your implementation case, btw. Regards HASSAN On Sat, Aug 21, 2010 at 22:55, David Ponzone <david.ponzone@ipeva.fr> wrote:

Guys,

As it seems that this debate is going to last forever without an acceptable conclusion, let's try to refocus on something positive.

What is interesting for most of us are the real-world performances. As I said previously, the performance reports page is ready: http://wiki.freeswitch.org/wiki/Real-world_results

It would be really nice if the people who have interesting figures could start filling it, as it would probably take them 5 to 10 minutes!

David Ponzone Direction Technique email: david.ponzone@ipeva.fr tel: 01 74 03 18 97 gsm: 06 66 98 76 34

Service Client IPeva tel: 0811 46 26 26 www.ipeva.fr - www.ipeva-studio.com

*Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. **IPeva** décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur.* * *

Le 21/08/2010 à 16:58, Vlasis Hatzistavrou (KTI) a écrit :

Hello Jean-Yves,

Thanks for the effort to calm me down, but I am calm and fine. :)

Just for the record I wasn't the one doing the tests, I was just defending the guy who asked the original question because I've been in his place and felt for him.

;)

Rgds, Vlasis.

Jean-Yves F. Barbier wrote:

Le Sat, 21 Aug 2010 15:41:13 +0300,

"Vlasis Hatzistavrou (KTI)" <vhatz@kinetix.gr> a écrit :

PLS Vlasis, calm down, I think eveyone is loosing temper riht now.

Let's say you took a test option, whether it is a good one or not could be

debated for a long time, but if YOU think it is the right way to go for

YOUR needs then go with it!

So my advice would be to eventually adjust your method only if you think

it would help in your particular case of operation; and stick to it from

now in order to always use the same test environment, to have a constant

relevance of results and thus constant comparisons possibilities too in

time :)

(nooo, not in the head!!!:)

JY

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

Nyamul Hassan wrote:

Vlasis, Woody's 2nd email talked about 600 CPS on v1.04. And, perhaps you can share your usage scenario, and how much pressure you could put on FreeSwitch before it "crashed"?

I must have missed Woody's 2nd email...

Will like to see your implementation case, btw.

Sure, I'll be happy to. Should I contact you off list? Best regards, Vlasis.

Nyamul Hassan wrote:

Why? On list is fine. Apparently many would like to see the "performance" that others get from their implementation.

Well, I thought too that many would like to see performance stats from other users, but after all the discussion in this thread I drew the conclusion that an on-list discussion about performance was actually _not_ fine. After all, you wrote that you understood the devs' irritation about performance discussions and this is why I proposed to send you the data off list, since you are interested.

You could also contribute to the "wiki page", that David was referring to.

I'll just proceed to post the data on the wiki page, then. Rgds, Vlasis.

Hello David, Apologies for the delay, it was a very busy day at work, it always on Mondays. I'll post the results asap, as I'll have to do some digging around in old files first. Best regards, Vlasis. On 23/8/2010 5:07 μμ, David Ponzone wrote:

Vlasis,

I don't see your results on the wiki page yet. I don't want to harass you but I am sure if some people start to fill it, others will follow. If we don't do that, I am afraid the table will still be empty in 1 year...

I would gladly put my own results but they are far from interesting, as we don't push much traffic through FS yet.

Com'on everyone! 5 minutes to add one line in the table! If you feel lazy or if you are wiki-proof, send me a private mail with the information, and I'll update it.

David Ponzone Direction Technique email: david.ponzone@ipeva.fr <mailto:david.ponzone@ipeva.fr> tel: 01 74 03 18 97 gsm: 06 66 98 76 34

Service Client IPeva tel: 0811 46 26 26 www.ipeva.fr <BLOCKED::http://www.ipeva.fr/> - www.ipeva-studio.com <BLOCKED::http://www.ipeva-studio.com/>

/Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. /*/IPeva/*/ décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur./ / /

On Fri, Aug 20, 2010 at 07:19:43PM +0300, Vlasis Hatzistavrou wrote:

On 20/8/2010 5:52 ??, Anthony Minessale wrote:

...

Yet another load test thread.........................

This is why we don't like them.

Too bad you missed my presentation at cluecon.....

Do you hear yourself btw? Quibbling about completely free software *only* doing 200cps in your fake test? I don't know if I missed any message in this thread, but why do you say that the tester did a fake test?

...

200cps is 4 times the industry standard FYI and if you were really doing 200cps in real life you would be so rich at our expense that you could afford more boxes.

you clearly ignored all of our repeated recommendations. Use the following or you are on your own.......

This OS Centos 5.x x86_64

If someone is on his own for any other OS than Centos 5.x x86_64 then why is FS offered for so many different platforms?

It's only natural for people to want to test it on their favorite platform... Perhaps, as you recommend, Centos 5.x x86_64 gives the best results over all other platforms, but last time I checked, a user is still allowed to ask performance related questions for his favorite platform in this mailing list, right?

I agree. Woody Dickson said What I found so strange was that while CPU usage remains pretty low and distributed among all cores at 190 - 200 calls per second. Then, after added a few more calls per second, all CPU becomes fully utilized. I did not interpret his comments as having anything to do with criticizing or questioning the performance of freeswitch. His question was one of curiosity about the extreme non-linearity of the processor load suddenly surging by adding just a few calls. If I had seen this same behaviour I likely would have asked the same question. Just my two cents worth. Vince

Woody's post was not expressing any negative opinion about FS's performance. He didn't comment on whether 200cps was too little or too much. He just wrote that the CPU usage increased in an unexpected manner and wonders how/if he can solve this. He wrote that he was doing a load test scenario, he didn't write that he was making money at 200cps "at your expense". And he wrote all this in a manner which was neither insulting nor abusive.

The poster just wanted to learn something, see if others have gotten better results with other setups. He was trying to make a valid discussion to solve a problem, and other users were actually replying to him.

This is what mailing lists are for, right?

I really don't understand why red flags have to be raised every time the word "performance" is mentioned in this mailing list.

-- Best regards, Vlasis Hatzistavrou.

Hi Vince, I have played with running Freeswitch on BSD too but the result is not great. The reason seems to be because BSD's threading is not as efficient as the one in Linux or there may be some other ways to tune it. BSD does give a better pure UDP throughput performance by the way. So what I ended up doing is developing my own UDP implementation which enable media to move through the ethernet at raw wire speed. I am able to max out the ethernet card limitation on Linux platform as a result of that. Woody On Mon, Aug 23, 2010 at 1:53 AM, Vincent Stemen <vince.freeswitch@hightek.org> wrote:

On Fri, Aug 20, 2010 at 05:29:31PM +0800, Woody Dickson wrote:

...

Hi,

I am doing some experiments with Freeswitch by torturing it to see how the machine's CPU response to heavy loaded situation. The test is done on a 16 core 5550 dual quad core server running fedora 2.6.30.10-105.2.23.fc11.x86_64 OS.

What I found so strange was that while CPU usage remains pretty low and distributed among all cores at 190 - 200 calls per second.  Then, after added a few more calls per second, all CPU becomes fully utilized.

Is this due to some wrong setting?  Any idea how I can tweak the configuration and continue my test?

Thanks, Woody

Hi Woody.

I would hazard to guess that this could be a Linux resource management issue.  I don't have any experience with SMP on Linux, but Linux has a long history of memory management (among other) problems.  We ran Linux exclusively on all our servers and workstations for over 10 years before finally switching to BSD several years ago.  We had continuous problems ranging from minor strange unexplained behaviours, as you describe, to what appeared to be bugs in applications, to outright crashes and freezes of the whole OS every day.  When we switched to BSD nearly all the problems went away.  Even some of the (what appeared to be) bugs in Linux binary applications went way, going from Linux to BSD running under Linux emulation (without re-compiling), using the same Linux libraries on the same hardware.

An interesting test would be to try the same load test with BSD on the same machine and see if you get a similar result.

Regards, Vince

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

Based on my experiment, that is still the case with CentOS. After using my own UDP, I was able to get 70K pps on 1 ethernet card on a Intel 5550. The limitation is that I don't have enough machine power to fire off enough calls to max it out. On Mon, Aug 23, 2010 at 8:25 PM, Nyamul Hassan <mnhassan@usa.net> wrote:

Is this also the case for the recommended CentOS / RHEL?  Do you still have to resort to having your own UDP implementation to max out Eth Card limitation? In the past, I have found a limitation on Linux, that the eth driver is single-threaded.  So, I couldn't push beyond 50K pps on a Intel Quad E5504 HP machine through 1 ethernet card. Regards HASSAN

On Mon, Aug 23, 2010 at 09:34, Woody Dickson <woodydickson@gmail.com> wrote:

...

Hi Vince,

I have played with running Freeswitch on BSD too but the result is not great.  The reason seems to be because BSD's threading is not as efficient as the one in Linux or there may be some other ways to tune it.  BSD does give a better pure UDP throughput performance by the way.

So what I ended up doing is developing my own UDP implementation which enable media to move through the ethernet at raw wire speed.  I am able to max out the ethernet card limitation on Linux platform as a result of that.

Woody

On Mon, Aug 23, 2010 at 1:53 AM, Vincent Stemen <vince.freeswitch@hightek.org> wrote:

...

On Fri, Aug 20, 2010 at 05:29:31PM +0800, Woody Dickson wrote:

...

Hi,

I am doing some experiments with Freeswitch by torturing it to see how the machine's CPU response to heavy loaded situation. The test is done on a 16 core 5550 dual quad core server running fedora 2.6.30.10-105.2.23.fc11.x86_64 OS.

What I found so strange was that while CPU usage remains pretty low and distributed among all cores at 190 - 200 calls per second.  Then, after added a few more calls per second, all CPU becomes fully utilized.

Is this due to some wrong setting?  Any idea how I can tweak the configuration and continue my test?

Thanks, Woody

Hi Woody.

I would hazard to guess that this could be a Linux resource management issue.  I don't have any experience with SMP on Linux, but Linux has a long history of memory management (among other) problems.  We ran Linux exclusively on all our servers and workstations for over 10 years before finally switching to BSD several years ago.  We had continuous problems ranging from minor strange unexplained behaviours, as you describe, to what appeared to be bugs in applications, to outright crashes and freezes of the whole OS every day.  When we switched to BSD nearly all the problems went away.  Even some of the (what appeared to be) bugs in Linux binary applications went way, going from Linux to BSD running under Linux emulation (without re-compiling), using the same Linux libraries on the same hardware.

An interesting test would be to try the same load test with BSD on the same machine and see if you get a similar result.

Regards, Vince

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

I just used normal g711. This involve some code changes in the freeswitch core as well. I need to package it a bit better and get some more experiments done to fully qualify it. The problem with this kind of tests is getting enough hardware to generate this amount of calls is tricky. On Mon, Aug 23, 2010 at 10:02 PM, David Ponzone <david.ponzone@ipeva.fr> wrote:

Woody, what codec do you use to push 70kpps on 1 GigE card ? David Ponzone  Direction Technique email: david.ponzone@ipeva.fr tel:      01 74 03 18 97 gsm:   06 66 98 76 34 Service Client IPeva tel:      0811 46 26 26 www.ipeva.fr  -   www.ipeva-studio.com Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur.

Le 23/08/2010 à 15:44, Woody Dickson a écrit :

Based on my experiment, that is still the case with CentOS.  After using my own UDP, I was able to get 70K pps on 1 ethernet card on a Intel 5550.  The limitation is that I don't have enough machine power to fire off enough calls to max it out.

On Mon, Aug 23, 2010 at 8:25 PM, Nyamul Hassan <mnhassan@usa.net> wrote:

Is this also the case for the recommended CentOS / RHEL?  Do you still have

to resort to having your own UDP implementation to max out Eth Card

limitation?

In the past, I have found a limitation on Linux, that the eth driver is

single-threaded.  So, I couldn't push beyond 50K pps on a Intel Quad E5504

HP machine through 1 ethernet card.

Regards

HASSAN

On Mon, Aug 23, 2010 at 09:34, Woody Dickson <woodydickson@gmail.com> wrote:

Hi Vince,

I have played with running Freeswitch on BSD too but the result is not

great.  The reason seems to be because BSD's threading is not as

efficient as the one in Linux or there may be some other ways to tune

it.  BSD does give a better pure UDP throughput performance by the

way.

So what I ended up doing is developing my own UDP implementation which

enable media to move through the ethernet at raw wire speed.  I am

able to max out the ethernet card limitation on Linux platform as a

result of that.

Woody

On Mon, Aug 23, 2010 at 1:53 AM, Vincent Stemen

<vince.freeswitch@hightek.org> wrote:

On Fri, Aug 20, 2010 at 05:29:31PM +0800, Woody Dickson wrote:

Hi,

I am doing some experiments with Freeswitch by torturing it to see how

the machine's CPU response to heavy loaded situation.

The test is done on a 16 core 5550 dual quad core server running

fedora 2.6.30.10-105.2.23.fc11.x86_64 OS.

What I found so strange was that while CPU usage remains pretty low

and distributed among all cores at 190 - 200 calls per second.  Then,

after added a few more calls per second, all CPU becomes fully

utilized.

Is this due to some wrong setting?  Any idea how I can tweak the

configuration and continue my test?

Thanks,

Woody

Hi Woody.

I would hazard to guess that this could be a Linux resource management

issue.  I don't have any experience with SMP on Linux, but Linux has

a long history of memory management (among other) problems.  We ran

Linux exclusively on all our servers and workstations for over 10 years

before finally switching to BSD several years ago.  We had continuous

problems ranging from minor strange unexplained behaviours, as you

describe, to what appeared to be bugs in applications, to outright

crashes and freezes of the whole OS every day.  When we switched to BSD

nearly all the problems went away.  Even some of the (what appeared to

be) bugs in Linux binary applications went way, going from Linux to BSD

running under Linux emulation (without re-compiling), using the same

Linux libraries on the same hardware.

An interesting test would be to try the same load test with BSD on the

same machine and see if you get a similar result.

Regards,

Vince

_______________________________________________

FreeSWITCH-users mailing list

FreeSWITCH-users@lists.freeswitch.org

http://lists.freeswitch.org/mailman/listinfo/freeswitch-users

UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users

http://www.freeswitch.org

_______________________________________________

FreeSWITCH-users mailing list

FreeSWITCH-users@lists.freeswitch.org

http://lists.freeswitch.org/mailman/listinfo/freeswitch-users

UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users

http://www.freeswitch.org

_______________________________________________

FreeSWITCH-users mailing list

FreeSWITCH-users@lists.freeswitch.org

http://lists.freeswitch.org/mailman/listinfo/freeswitch-users

UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users

http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

Chances are you're not hitting the bandwidth limits but the fact that you're moving tiny packets around. Smaller packets == more context switches == less throughput. You can increase your packet size to 60ms and gain performance. /b On Aug 23, 2010, at 9:38 AM, David Ponzone wrote:

So the throughput on the card topped at around 150Mbps, if I calculate correctly ?

David Ponzone Direction Technique email: david.ponzone@ipeva.fr tel: 01 74 03 18 97 gsm: 06 66 98 76 34

Service Client IPeva tel: 0811 46 26 26 www.ipeva.fr - www.ipeva-studio.com

Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur.

I think you explained it a bit more in detail... we are saying exactly the same thing just I'm a bit more brief. /b On Aug 23, 2010, at 9:47 AM, Nyamul Hassan wrote:

David,

In our experience, it is not about "bps", but about "pps". Because the ethernet driver in Linux is single threaded, you cannot push beyond what a single core can handle.

In our case, after 60kpps (on Intel X5504), the quality deteriorated. In Woody's case, this limit goes as high as 70kpps (on Intel X5550), which is roughly equivalent to the single core clock speed difference between the servers.

Brian is speaking the exact same jargon, that I came across when I was reading about this 1 year ago.

Regards HASSAN

2010/8/23 Brian West <brian@freeswitch.org> Chances are you're not hitting the bandwidth limits but the fact that you're moving tiny packets around. Smaller packets == more context switches == less throughput. You can increase your packet size to 60ms and gain performance.

/b

It's Always this type of thread that gets attention an not others that actually need attention. I may not have been cute and cuddly in my reply to Woody but I gave him a series of load testing aids in that same reply. And I don't notice Woody complaining. If you are offended Woody, accept my regrets. You have indeed hit a sore note because this subject always creates a huge pain for us and I don't always have time for tact in my busy day. Did you receive the numerous tips I suggested in my replies? I'll restate the most important which is probably to use the -heavy-timer command line arg to FS on a box like yours. (This was default in 1.0.4) I believe most of what I said is already on our WIKI but if not it would be a good idea to get it in there. READERS: Please strike "at our expense" from my original reply and replace it with "with the help of our absolutely free software" as some of the finer nuances in our irregular english language especially idioms may be lost in translation for some individuals. To the rest of you who have stepped up to make a non-biased performance WIKI page, Thank you for your efforts perhaps that will save us all some time in the long run. On Mon, Aug 23, 2010 at 10:34 AM, David Ponzone <david.ponzone@ipeva.fr> wrote:

Thank you guys, but I should have said that I am perfectly aware of the pps vs. bps limitations. I had several years of experience dealing with Cisco 7200/7500 to learn about how one vendor can lie about performances :) I was just asking this to be sure about the figures. 70k pps mean around 1400 concurrent calls if ptime is 20ms (I am assuming 70k pps means 70k for both ways. If it's 70k pps with input/output aggregated, all the following figures have to be divided by 2). I heard that some people around get 3000 calls on one box. So they get those 3000 calls probably spreading the load among multiple gigE cards. It's an important information to have for people building quite large switches that one gigE card won't handle more than 1000 calls with a regular kernel. Perhaps even 800 if you want to stay on the safe side. David Ponzone  Direction Technique email: david.ponzone@ipeva.fr tel:      01 74 03 18 97 gsm:   06 66 98 76 34 Service Client IPeva tel:      0811 46 26 26 www.ipeva.fr  -   www.ipeva-studio.com Ce message et toutes les pièces jointes sont confidentiels et établis à l'intention exclusive de ses destinataires. Toute utilisation ou diffusion non autorisée est interdite. Tout message électronique est susceptible d'altération. IPeva décline toute responsabilité au titre de ce message s'il a été altéré, déformé ou falsifié. Si vous n'êtes pas destinataire de ce message, merci de le détruire immédiatement et d'avertir l'expéditeur.

Le 23/08/2010 à 17:17, Brian West a écrit :

I think you explained it a bit more in detail... we are saying exactly the same thing just I'm a bit more brief.

/b

On Aug 23, 2010, at 9:47 AM, Nyamul Hassan wrote:

David,

In our experience, it is not about "bps", but about "pps".  Because the ethernet driver in Linux is single threaded, you cannot push beyond what a single core can handle.

In our case, after 60kpps (on Intel X5504), the quality deteriorated.  In Woody's case, this limit goes as high as 70kpps (on Intel X5550), which is roughly equivalent to the single core clock speed difference between the servers.

Brian is speaking the exact same jargon, that I came across when I was reading about this 1 year ago.

Regards

HASSAN

2010/8/23 Brian West <brian@freeswitch.org>

Chances are you're not hitting the bandwidth limits but the fact that you're moving tiny packets around.  Smaller packets == more context switches == less throughput.  You can increase your packet size to 60ms and gain performance.

/b

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

_______________________________________________ FreeSWITCH-users mailing list FreeSWITCH-users@lists.freeswitch.org http://lists.freeswitch.org/mailman/listinfo/freeswitch-users UNSUBSCRIBE:http://lists.freeswitch.org/mailman/options/freeswitch-users http://www.freeswitch.org

-- Anthony Minessale II FreeSWITCH http://www.freeswitch.org/ ClueCon http://www.cluecon.com/ Twitter: http://twitter.com/FreeSWITCH_wire AIM: anthm MSN:anthony_minessale@hotmail.com GTALK/JABBER/PAYPAL:anthony.minessale@gmail.com IRC: irc.freenode.net #freeswitch FreeSWITCH Developer Conference sip:888@conference.freeswitch.org googletalk:conf+888@conference.freeswitch.org pstn:+19193869900