Re: Converged Load Balancer

[Ivelin Ivanov]

This is great info. That quite a broad range of flexibility depending
on the application usage patterns.
I'm copying mobicents-public. Would be cool if you have time for a
blog piece with examples for the main scenarios.

Ivelin


On Wed, Nov 18, 2009 at 3:24 PM, Vladimir Ralev
<vladimi...@gmail.com> wrote:
> Since the integrated HTTP balancer requires zero extra configuration
> it is always best to start with it.
>
> We definitely recommend single LB for HTTP and SIP when they don't
> have high perf requirements say HTTP + SIP < 3000rps. Also after
> failure of a node the integrated HTTP balancer takes more time to
> switch (depends on the heartbeat, while mod_jk uses ICMP or similar
> methods). So if the users can live with these limitations, which I
> guess would be most of the users, they are fine.
>
> When the HTTP requests are a lot more and outnumber the SIP requests
> greatly, it's better to use apache for HTTP. It is completely possible
> to use apache and the integrated LB together in the same deployment.
> For example if most users do not use telco features they will browse
> telco-free part of the website through apache (company products,
> forums, etc), but when they start to use telco feature and need to see
> conference status for example they can be redirected to the integrated
> HTTP forwarder where affinity keys are analyzed and will be grouped
> with SIP sessions to stick to one node.
>
> A single integrated HTTP balancer has 3-5 x worse performance than
> mod_jk (AJP and native code make big difference under load). It is
> scalable and distributable on the other side, but my guess it users
> would not want to dedicate more than 2 machines for load-balancing.
>
> On Wed, Nov 18, 2009 at 10:51 PM, Ivelin Ivanov
> <ivelin.atan...@gmail.com> wrote:
>> Vladimir,
>>
>> Can you please tell us a bit more about the achievements. I wasn't
>> able to stick past 12 to ask questions. From the meeting notes, I
>> have:
>> - single balancer on latop show 600-700 cps for SIP UAS,
>> - 300-900 jboss AS homepage loads for integrated HTTP forwarder (AS
>> CPU utilization limits the test)
>> - comet HTTP works flawlessly (this very chat  app was used in the test)
>> - The only lost calls are caused by some probems in Sip Servlets with
>> retransmissions, but those are now eliminated.
>>
>> Does this mean that we can now recommend a single LB for both HTTP and SIP?
>>
>> Ivelin
>>
>> --

[Vladimir Ralev]

Yeah, I am updating the slides regularly and they are on my blog. The
biggest problem is to fit the large diagrams needed to visualize it
better. I will figure it out. Just a clarification for everyone - the
Mobicents Load balancer operates with pluggable algorithms - you can
choose a built-in or write your own. Thus performance depends on the
algorithm. Also keep in mind that the apache comparison perf number
are with limited accuracy and the numbers are probably a bit better in
reality with the defaults and can be optimized with
application-specific algorithm. Would love if someone from the
community wants to check it out.

On Wed, Nov 18, 2009 at 11:57 PM, Ivelin Ivanov

[Vladimir Ralev]

http://vladimirralev.blogspot.com/2009/10/mobicents-sip-load-balancing.html

But I will write a dedicated blogpost as testing goes on.

[Ivelin Ivanov]

Smoking hot stuff!

On Wed, Nov 18, 2009 at 10:00 PM, Vladimir Ralev