Icecast - Mar 2024 - Education - 1, 000s, 100, 000's, Millions of listeners. (What kind of infrastructure)

Dear all

 

My 5 cents (or Rappen in CH) if it comes to serving many clients.

We are running a 4 node cluster since several years ? rock solid and w/o any
issues. This cluster serves many thousands of listeners from all over the world.
Our source transcoder sending the audio streams to each Node.  Hence,
transcoding power is not an issue here. The four Nodes a geographically
dispersed in 3 countries in Europe. In our case each Node is running Debian with
icecast and has 10Gbit connectivity with brilliant worldwide peerings. Good
peering is key, choose your ISP wisely ?.Each icecast servers has the same multi
domain ssl cert. which allows us to deliver to several customers (each customer
a subdomain) the cluster is round robin load balanced by using AWS Route53. This
approach may can be achived also with other DNS Providers like Cloudflare. For
example, if one node need to be taken down for maintenance, Route53 throws the
Node out of the DNS automatically. This will be achived with ?health checks?
This mechanism is pretty fast and responsive. If a client gets disconnected and
tries a reconnect, the RR DNS is passing the client immediately to a working
Node. No issues here as well. It just works.

 

In the beginning we?ve experienced similar issues even though, the bandwidth
capacity of the VM?s was never the root cause. We?ve identified some solutions:

 

1.	Linux TCP stack tuning. Cloudflare has many studies published in their Blogs
about this. But you will find a lot about this tuning also in the interweb
2.	Consider to bake your own kernel, which is tuned for high throughput ? goes
in line with TCP Stack tunning
3.	Tune the Linux open file limits and adjust the init start script for the
icecast server. Example : start icecast with  ulimit -c unlimited and ulimit -n
32768
4.	Consider to use FreeBSD instead of Linux. FreeBSD has the better TCP stack
out of the box.
5.	If all of this is not feasible for you, just add a new Node to the cluster
and level the amount of clients to more Nodes.

 

For the points 1. and 2. I can?t give you a ?out of the box? solution or default
settings. It?s an iteration process: adjusting, trial, load testing, monitoring
and <repeat>

Because the result will need to fit your requirements, therefore every setup
might need a different tuning. And btw. do not try using icecast on Windows
Servers, if you need to serve a lot of clients ?

 

Happy icecasting

tom

 

Von: Icecast <icecast-bounces at xiph.org> Im Auftrag von Fred Gleason
Gesendet: Mittwoch, 20. M?rz 2024 20:53
An: Icecast streaming server user discussions <icecast at xiph.org>
Betreff: Re: [Icecast] Education - 1, 000s, 100, 000's, Millions of
listeners. (What kind of infrastructure)

 

On Mar 20, 2024, at 13:16, Wayne Barron <wayne at cffcs.com <mailto:wayne
at cffcs.com> > wrote:

 

In Windows and Linux web servers, we can create a forest for our web servers.
Send traffic to different servers to even the workload.

Can we do something like this with the Icecast servers?
(or)
Will we have to install new VMs, add the heavy stations on that one,
and send the new traffic there?

 

Ok, I?m going to be ?that guy??

 

I would argue that, as soon as you?ve hit an audience size of 10,000 or more
(especially if that audience is at all geographically dispersed), IceCast is
basically off the table. The reason why can be summarized in three letters:
?CDN? [Content Distribution Networks].

 

To fan out to large, geographically dispersed audiences of 10,000 or more (not
to mention 100k?s or, Lord help us, 1M?s or more), you need to get content
cached in locations that are geographically close to your listeners. By far the
easiest (read: most cost effective) way to do this at scale is to leverage the
already existing infrastructure of CDNs (companies like Akamai or CloudFlare,
that have a world-wide footprint). That means using streaming formats that
utilize segmented distribution mechanisms, such as HLS or DASH. You can
kinda-sorta do this sort of thing with IceCast by using relays, but it?s complex
to configure and monitor while being not well supported at many CDNs (Akamai for
example discontinued their IceCast product offering several years ago). HLS OTOH
plays very well with that infrastructure because it?s effectively just a bunch
of static files that get replicated via HTTP[S]. No special ?server? software is
required; bog-standard Apache or Nginx work just fine, because the complex
?media handling? bits have been intentionally pushed to the endpoints; namely
the encoder and (especially) the players. Today though, when FOSS HLS audio
encoders are available and pretty much every browser supports playing HLS
content natively, the complexity angle can be largely ignored by content
creators.

 

Just my take. That, and 2 ? will get you a (cheap) cup of coffee?

 

Cheers!

 

 

|---------------------------------------------------------------------|

| Frederick F. Gleason, Jr. |             Chief Developer             |

|                           |             Paravel Systems             |

|---------------------------------------------------------------------|

|    All progress is based upon a universal innate desire of every    |

|    organism to live beyond its income.                              |

|                                                                     |

|                                                  -- Samuel Butler   |

|                                                  "Notebooks"       
|

|---------------------------------------------------------------------|

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.xiph.org/pipermail/icecast/attachments/20240320/d626b96f/attachment.htm>

Tom and Frederick.
Thank you both for your input. It is greatly appreciated, not only by me,
but I am sure for many others who find this thread.

Tom - using Linux server for icecast.

HLS was brought to my attention a while back on the liquidsoap forum.
I have not had a chance to completely look in on it, but do plan on
checking it out.

Tom. You have your icecast servers in a round robin configuration.
Are you using NGINX for that?
I have watched several YouTube videos on using it for doing round robin as
well as ssl and other things.

I understand that once a scenario of say 10,000 or more, that I will have
to look into either a bigger infrastructure on my side or leasing space on
the cloud somewhere.

Now, if I did do the cloud.
Would that be to host the servers or just the files or everything?

Wayne


On Wed, Mar 20, 2024, 4:50?PM <thomas.zumbrunnen at gmail.com> wrote:
> Dear all
>
>
>
> My 5 cents (or Rappen in CH) if it comes to serving many clients.
>
> We are running a 4 node cluster since several years ? rock solid and w/o
> any issues. This cluster serves many thousands of listeners from all over
> the world. Our source transcoder sending the audio streams to each Node.
> Hence, transcoding power is not an issue here. The four Nodes a
> geographically dispersed in 3 countries in Europe. In our case each Node
> is running Debian with icecast and has 10Gbit connectivity with brilliant
> worldwide peerings. Good peering is key, choose your ISP wisely ?.Each
> icecast servers has the same multi domain ssl cert. which allows us to
> deliver to several customers (each customer a subdomain) the cluster is
> round robin load balanced by using AWS Route53. This approach may can be
> achived also with other DNS Providers like Cloudflare. For example, if one
> node need to be taken down for maintenance, Route53 throws the Node out of
> the DNS automatically. This will be achived with ?health checks? This
> mechanism is pretty fast and responsive. If a client gets disconnected and
> tries a reconnect, the RR DNS is passing the client immediately to a
> working Node. No issues here as well. It just works.
>
>
>
> In the beginning we?ve experienced similar issues even though, the
> bandwidth capacity of the VM?s was never the root cause. We?ve identified
> some solutions:
>
>
>
>    1. Linux TCP stack tuning. Cloudflare has many studies published in
>    their Blogs about this. But you will find a lot about this tuning also
in
>    the interweb
>    2. Consider to bake your own kernel, which is tuned for high
>    throughput ? goes in line with TCP Stack tunning
>    3. Tune the Linux open file limits and adjust the init start script
>    for the icecast server. Example : start icecast with  ulimit -c
unlimited
>    and ulimit -n 32768
>    4. Consider to use FreeBSD instead of Linux. FreeBSD has the better
>    TCP stack out of the box.
>    5. If all of this is not feasible for you, just add a new Node to the
>    cluster and level the amount of clients to more Nodes.
>
>
>
> For the points 1. and 2. I can?t give you a ?out of the box? solution or
> default settings. It?s an iteration process: adjusting, trial, load
> testing, monitoring and <repeat>
>
> Because the result will need to fit your requirements, therefore every
> setup might need a different tuning. And btw. do not try using icecast on
> Windows Servers, if you need to serve a lot of clients ?
>
>
>
> Happy icecasting
>
> tom
>
>
>
> *Von:* Icecast <icecast-bounces at xiph.org> *Im Auftrag von *Fred
Gleason
> *Gesendet:* Mittwoch, 20. M?rz 2024 20:53
> *An:* Icecast streaming server user discussions <icecast at xiph.org>
> *Betreff:* Re: [Icecast] Education - 1, 000s, 100, 000's, Millions of
> listeners. (What kind of infrastructure)
>
>
>
> On Mar 20, 2024, at 13:16, Wayne Barron <wayne at cffcs.com> wrote:
>
>
>
> In Windows and Linux web servers, we can create a forest for our web
> servers.
> Send traffic to different servers to even the workload.
>
> Can we do something like this with the Icecast servers?
> (or)
> Will we have to install new VMs, add the heavy stations on that one,
> and send the new traffic there?
>
>
>
> Ok, I?m going to be ?that guy??
>
>
>
> I would argue that, as soon as you?ve hit an audience size of 10,000 or
> more (especially if that audience is at all geographically dispersed),
> IceCast is basically off the table. The reason why can be summarized in
> three letters: ?CDN? [Content Distribution Networks].
>
>
>
> To fan out to large, geographically dispersed audiences of 10,000 or more
> (not to mention 100k?s or, Lord help us, 1M?s or more), you need to get
> content cached in locations that are geographically close to your
> listeners. By far the easiest (read: most cost effective) way to do this at
> scale is to leverage the already existing infrastructure of CDNs (companies
> like Akamai or CloudFlare, that have a world-wide footprint). That means
> using streaming formats that utilize segmented distribution mechanisms,
> such as HLS or DASH. You can kinda-sorta do this sort of thing with IceCast
> by using relays, but it?s complex to configure and monitor while being not
> well supported at many CDNs (Akamai for example discontinued their IceCast
> product offering several years ago). HLS OTOH plays very well with that
> infrastructure because it?s effectively just a bunch of static files that
> get replicated via HTTP[S]. No special ?server? software is required;
> bog-standard Apache or Nginx work just fine, because the complex ?media
> handling? bits have been intentionally pushed to the endpoints; namely the
> encoder and (especially) the players. Today though, when FOSS HLS audio
> encoders are available and pretty much every browser supports playing HLS
> content natively, the complexity angle can be largely ignored by content
> creators.
>
>
>
> Just my take. That, and 2 ? will get you a (cheap) cup of coffee?
>
>
>
> Cheers!
>
>
>
>
>
> |---------------------------------------------------------------------|
>
> | Frederick F. Gleason, Jr. |             Chief Developer             |
>
> |                           |             Paravel Systems             |
>
> |---------------------------------------------------------------------|
>
> |    All progress is based upon a universal innate desire of every    |
>
> |    organism to live beyond its income.                              |
>
> |                                                                     |
>
> |                                                  -- Samuel Butler   |
>
> |                                                  "Notebooks"   
|
>
> |---------------------------------------------------------------------|
>
>
> _______________________________________________
> Icecast mailing list
> Icecast at xiph.org
> http://lists.xiph.org/mailman/listinfo/icecast
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.xiph.org/pipermail/icecast/attachments/20240320/187fc23b/attachment.htm>