short update on performance and Red5

[Walter Tak]

Some parts of this message have been removed. Learn more about Nabble's security policy.

For those that are a little interested in the performance of Red5 and VOD (video on demand e.g. no live-streams) then here are some measured timings:

 

I executed a few times 10 threads that each downloads as fast as possible 1 single VOD file ( 60 Mb ).

The downloading client is a quad-core xeon @ 2.13 ghz (X3210) on a 100 mbit link at a datacentre.

 

The Red5-server serving the files has the same configuration and had a load of 17% CPU (or 75% on one CPU) during the process while outputting at maximum network-capable speed (e.g. 100 mbit , 11 Mbytes). Before and after the process the cpu dropped to a around 0%.

 

10 threads each downloading a different VOD file (each between 30-60 Mb each) didn't differ at all ; same speed (max network) and same load.

 

Based on that very short test one could say that such a config could serve around 5 times that output (e.g. 500 mbit) before maxing out with CPU.

 

With a faster CPU (newer Xeons) that run around 3 Ghz and faster FSB's that value might even be increased with 50 to 100%. You might hit 750-900 Mbit per server,

in theory. 100 mbit continous streaming with a few streams didn't make my server sweat.

 

So if you want to calculate "how many ppl can watch my videos" then just divide bandwidth by bitrate. If your box is powerful enough (simple dual core should be able to handle 100 mbit without problems) and you have the 100 mbit really dedicated uplink then you could serve 100 streams of 1 mbit each, or 1000 streams of 100 kbit each. (And of mix ofcourse).

 

With a real 1 gbit connection (make sure you have someone paying the bills for that) you could increase that per box with a factor 5-8.

 

With 1 mbit you can show decent H.264 content, with a faster server, 1 gbit, you could do around 500-800 per box. Let's assume 600.

 

Someone asked for 2 million viewers (right?) ; if those were simultaneous you'd need 2.000.000 / 800 = 2500 servers. Each pumping out a massive 800-900 Mbit.

No need to say that's quite an enterprise job. If the viewers aren't simultaneous but just 2 million visitors per month, each viewing one video then you can

count how many servers you need with bitrate and length of the video.

 

E.g. 800 ppl can watch simultaneous on your server. Your videos have an average length of 10 minutes (youtube concept). You can serve 800 users * ( 60 minutes per hour / 10 minutes length) = 4800 videos per hour. If you visitors would be spread out over 10 hours of a day (instead of 24 hours a day) you could serve 4800 * 10 = 48.000 videos per day. Let's reserve some days and say 25 active days per month and your server would do 48.000 * 25 = 1.2 million video-views each month.

 

 

 

My 2 cents,

 

Walter

 

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

[Walter Tak]

Some parts of this message have been removed. Learn more about Nabble's security policy.

Changes in Italic since original message was a bit sloppy.

 

For those that are a little interested in the performance of Red5 and VOD (video on demand e.g. no live-streams) then here are some measured timings:

 

I executed a few times 10 threads that each downloads as fast as possible 1 single VOD file ( 60 Mb ) via RTMP.

The downloading client is a dedicated quad-core xeon @ 2.13 ghz (X3210) on a 100 mbit link , hosted in a datacentre.

 

The Red5-server serving the files has the same configuration and had a load of 17% CPU (or 75% on one CPU) during the process while outputting at maximum network-capable speed (e.g. 100 mbit , 11 Mbytes). Before and after the process the cpu dropped to a around 0%.

 

10 threads each downloading a different VOD file (each between 30-60 Mb each) didn't differ at all ; same speed (max network) and same load.

 

Based on that very short test one could say that the same configuration , with a 1 gbit NIC instead of 100 mbit, could serve around 5 times that output (e.g. 500 mbit) before maxing out with CPU.

 

With a faster CPU (newer Xeons) that run around 3 Ghz and faster FSB's that value might even be increased with 50 to 100%. You might hit 750-900 Mbit per server,

in theory. 100 mbit continous streaming with a few streams didn't make my server sweat for sure.

 

So if you want to calculate "how many ppl can watch my videos before my servers are maxxed out" then just divide bandwidth by bitrate. If your box is powerful enough (simple dual core should be able to handle 100 mbit without problems) and you have the 100 mbit really dedicated uplink then you could serve 100 streams of 1 mbit each, or 1000 streams of 100 kbit each. (Or a mix ofcourse).

 

With a real 1 gbit connection (make sure you have someone paying the bills for that) you could increase that per box with a factor 5-8.

 

With a video of 1 mbit bitrate you can show decent H.264 content,near fullscreen, with a faster server and a 1 gbit NIC , you could do around 500-800 per box. Let's assume 800 VODs, each 1mbit, fast quad-xeon with a 1gbit NIC.

 

Someone asked for 2 million viewers (right?) ; if those were simultaneous you'd need 2.000.000 / 800 = 2500 servers. Each pumping out a massive 800-900 Mbit. Pricing for 1gbit flatfee bandwidth could approach 500-2000 usd per server so you're talking about 1.2 to 5 million USD a month. Perhaps you can make a better deal with that many servers ofcourse.

 

No need to say that's quite an enterprise job. If the viewers aren't simultaneous but just 2 million visitors per month, each viewing one video then you can

count how many servers you need with bitrate and length of the video.

 

E.g. 800 ppl can watch simultaneous on your server. Your videos have an average length of 10 minutes (youtube concept). You can serve 800 users * ( 60 minutes per hour / 10 minutes length) = 4800 videos per hour. If you visitors would be spread out over 10 hours of a day (instead of 24 hours a day) you could serve 4800 * 10 = 48.000 videos per day. Let's reserve some days and say 25 active days per month and your server would do 48.000 * 25 = 1.2 million video-views each month.

 

You could increase the maximum amount of simultaneous viewers by adding multiple 1 gbit nics in one server but even a fast quad core CPU will be your bottleneck.

 

My 2 cents,

 

Walter

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

[Argenacho]

Hey Walter, thanks for sharing this analysis. Do you think this only applies to to VOD and not to live  streams for any particular reason? Shouldn't we expect a similar result?

Txs

On Thu, Jul 2, 2009 at 8:35 PM, Walter Tak <walter@...> wrote:

Changes in Italic since original message was a bit sloppy.

 

For those that are a little interested in the performance of Red5 and VOD (video on demand e.g. no live-streams) then here are some measured timings:

 

I executed a few times 10 threads that each downloads as fast as possible 1 single VOD file ( 60 Mb ) via RTMP.

The downloading client is a dedicated quad-core xeon @ 2.13 ghz (X3210) on a 100 mbit link , hosted in a datacentre.

 

The Red5-server serving the files has the same configuration and had a load of 17% CPU (or 75% on one CPU) during the process while outputting at maximum network-capable speed (e.g. 100 mbit , 11 Mbytes). Before and after the process the cpu dropped to a around 0%.

 

10 threads each downloading a different VOD file (each between 30-60 Mb each) didn't differ at all ; same speed (max network) and same load.

 

Based on that very short test one could say that the same configuration , with a 1 gbit NIC instead of 100 mbit, could serve around 5 times that output (e.g. 500 mbit) before maxing out with CPU.

 

With a faster CPU (newer Xeons) that run around 3 Ghz and faster FSB's that value might even be increased with 50 to 100%. You might hit 750-900 Mbit per server,

in theory. 100 mbit continous streaming with a few streams didn't make my server sweat for sure.

 

So if you want to calculate "how many ppl can watch my videos before my servers are maxxed out" then just divide bandwidth by bitrate. If your box is powerful enough (simple dual core should be able to handle 100 mbit without problems) and you have the 100 mbit really dedicated uplink then you could serve 100 streams of 1 mbit each, or 1000 streams of 100 kbit each. (Or a mix ofcourse).

 

With a real 1 gbit connection (make sure you have someone paying the bills for that) you could increase that per box with a factor 5-8.

 

With a video of 1 mbit bitrate you can show decent H.264 content,near fullscreen, with a faster server and a 1 gbit NIC , you could do around 500-800 per box. Let's assume 800 VODs, each 1mbit, fast quad-xeon with a 1gbit NIC.

 

Someone asked for 2 million viewers (right?) ; if those were simultaneous you'd need 2.000.000 / 800 = 2500 servers. Each pumping out a massive 800-900 Mbit. Pricing for 1gbit flatfee bandwidth could approach 500-2000 usd per server so you're talking about 1.2 to 5 million USD a month. Perhaps you can make a better deal with that many servers ofcourse.

 

No need to say that's quite an enterprise job. If the viewers aren't simultaneous but just 2 million visitors per month, each viewing one video then you can

count how many servers you need with bitrate and length of the video.

 

E.g. 800 ppl can watch simultaneous on your server. Your videos have an average length of 10 minutes (youtube concept). You can serve 800 users * ( 60 minutes per hour / 10 minutes length) = 4800 videos per hour. If you visitors would be spread out over 10 hours of a day (instead of 24 hours a day) you could serve 4800 * 10 = 48.000 videos per day. Let's reserve some days and say 25 active days per month and your server would do 48.000 * 25 = 1.2 million video-views each month.

 

You could increase the maximum amount of simultaneous viewers by adding multiple 1 gbit nics in one server but even a fast quad core CPU will be your bottleneck.

 

My 2 cents,

 

Walter

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

[Walter Tak]

Some parts of this message have been removed. Learn more about Nabble's security policy.

Since I didn't test it I don't want to make wrong assumptions.

 

More tests coming up this weekend perhaps.

 

In theory Red5 should have less problems with live-streams since it only needs to buffer a few packets in-memory in contrast to VODs which need to be read from disc.

 

W.

 

Hey Walter, thanks for sharing this analysis. Do you think this only applies to to VOD and not to live  streams for any particular reason? Shouldn't we expect a similar result?

Txs

On Thu, Jul 2, 2009 at 8:35 PM, Walter Tak <walter@...> wrote:

Changes in Italic since original message was a bit sloppy.

 

For those that are a little interested in the performance of Red5 and VOD (video on demand e.g. no live-streams) then here are some measured timings:

 

I executed a few times 10 threads that each downloads as fast as possible 1 single VOD file ( 60 Mb ) via RTMP.

The downloading client is a dedicated quad-core xeon @ 2.13 ghz (X3210) on a 100 mbit link , hosted in a datacentre.

 

The Red5-server serving the files has the same configuration and had a load of 17% CPU (or 75% on one CPU) during the process while outputting at maximum network-capable speed (e.g. 100 mbit , 11 Mbytes). Before and after the process the cpu dropped to a around 0%.

 

10 threads each downloading a different VOD file (each between 30-60 Mb each) didn't differ at all ; same speed (max network) and same load.

 

Based on that very short test one could say that the same configuration , with a 1 gbit NIC instead of 100 mbit, could serve around 5 times that output (e.g. 500 mbit) before maxing out with CPU.

 

With a faster CPU (newer Xeons) that run around 3 Ghz and faster FSB's that value might even be increased with 50 to 100%. You might hit 750-900 Mbit per server,

in theory. 100 mbit continous streaming with a few streams didn't make my server sweat for sure.

 

So if you want to calculate "how many ppl can watch my videos before my servers are maxxed out" then just divide bandwidth by bitrate. If your box is powerful enough (simple dual core should be able to handle 100 mbit without problems) and you have the 100 mbit really dedicated uplink then you could serve 100 streams of 1 mbit each, or 1000 streams of 100 kbit each. (Or a mix ofcourse).

 

With a real 1 gbit connection (make sure you have someone paying the bills for that) you could increase that per box with a factor 5-8.

 

With a video of 1 mbit bitrate you can show decent H.264 content,near fullscreen, with a faster server and a 1 gbit NIC , you could do around 500-800 per box. Let's assume 800 VODs, each 1mbit, fast quad-xeon with a 1gbit NIC.

 

Someone asked for 2 million viewers (right?) ; if those were simultaneous you'd need 2.000.000 / 800 = 2500 servers. Each pumping out a massive 800-900 Mbit. Pricing for 1gbit flatfee bandwidth could approach 500-2000 usd per server so you're talking about 1.2 to 5 million USD a month. Perhaps you can make a better deal with that many servers ofcourse.

 

No need to say that's quite an enterprise job. If the viewers aren't simultaneous but just 2 million visitors per month, each viewing one video then you can

count how many servers you need with bitrate and length of the video.

 

E.g. 800 ppl can watch simultaneous on your server. Your videos have an average length of 10 minutes (youtube concept). You can serve 800 users * ( 60 minutes per hour / 10 minutes length) = 4800 videos per hour. If you visitors would be spread out over 10 hours of a day (instead of 24 hours a day) you could serve 4800 * 10 = 48.000 videos per day. Let's reserve some days and say 25 active days per month and your server would do 48.000 * 25 = 1.2 million video-views each month.

 

You could increase the maximum amount of simultaneous viewers by adding multiple 1 gbit nics in one server but even a fast quad core CPU will be your bottleneck.

 

My 2 cents,

 

Walter

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

---

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

---

Internal Virus Database is out-of-date.
Checked by AVG.
Version: 7.5.560 / Virus Database: 270.12.26/2116 - Release Date: 15-05-09 06:16

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

[Andy Shaules]

Some parts of this message have been removed. Learn more about Nabble's security policy.

Live stream: power horse server with heap-big-pipe.

average 600 kbs  @ 350 viewers maximum.

 

My recommended high-quality, long-term, high-demand, no-crashes, 'Ican go to sleep' reliability setting.

 

More viewers than that, I burn out my drive belt because high gear is taller than the default settings can handle.(rtmp cuhunk size, and thread pool limits).

 

Can it handle more? yes...... Would I personally sell a single red5 server instance at a higher capacity capability which is used for HIGH quality streams? No... I would sugest distibuted deployment. 

 

:)

 

----- Original Message -----

From: walter@...

To: red5@...

Sent: Thursday, July 02, 2009 5:03 PM

Subject: Re: [Red5] short update on performance and Red5

Since I didn't test it I don't want to make wrong assumptions.

 

More tests coming up this weekend perhaps.

 

In theory Red5 should have less problems with live-streams since it only needs to buffer a few packets in-memory in contrast to VODs which need to be read from disc.

 

W.

 

Hey Walter, thanks for sharing this analysis. Do you think this only applies to to VOD and not to live  streams for any particular reason? Shouldn't we expect a similar result?

Txs

On Thu, Jul 2, 2009 at 8:35 PM, Walter Tak <walter@...> wrote:

Changes in Italic since original message was a bit sloppy.

 

For those that are a little interested in the performance of Red5 and VOD (video on demand e.g. no live-streams) then here are some measured timings:

 

I executed a few times 10 threads that each downloads as fast as possible 1 single VOD file ( 60 Mb ) via RTMP.

The downloading client is a dedicated quad-core xeon @ 2.13 ghz (X3210) on a 100 mbit link , hosted in a datacentre.

 

The Red5-server serving the files has the same configuration and had a load of 17% CPU (or 75% on one CPU) during the process while outputting at maximum network-capable speed (e.g. 100 mbit , 11 Mbytes). Before and after the process the cpu dropped to a around 0%.

 

10 threads each downloading a different VOD file (each between 30-60 Mb each) didn't differ at all ; same speed (max network) and same load.

 

Based on that very short test one could say that the same configuration , with a 1 gbit NIC instead of 100 mbit, could serve around 5 times that output (e.g. 500 mbit) before maxing out with CPU.

 

With a faster CPU (newer Xeons) that run around 3 Ghz and faster FSB's that value might even be increased with 50 to 100%. You might hit 750-900 Mbit per server,

in theory. 100 mbit continous streaming with a few streams didn't make my server sweat for sure.

 

So if you want to calculate "how many ppl can watch my videos before my servers are maxxed out" then just divide bandwidth by bitrate. If your box is powerful enough (simple dual core should be able to handle 100 mbit without problems) and you have the 100 mbit really dedicated uplink then you could serve 100 streams of 1 mbit each, or 1000 streams of 100 kbit each. (Or a mix ofcourse).

 

With a real 1 gbit connection (make sure you have someone paying the bills for that) you could increase that per box with a factor 5-8.

 

With a video of 1 mbit bitrate you can show decent H.264 content,near fullscreen, with a faster server and a 1 gbit NIC , you could do around 500-800 per box. Let's assume 800 VODs, each 1mbit, fast quad-xeon with a 1gbit NIC.

 

Someone asked for 2 million viewers (right?) ; if those were simultaneous you'd need 2.000.000 / 800 = 2500 servers. Each pumping out a massive 800-900 Mbit. Pricing for 1gbit flatfee bandwidth could approach 500-2000 usd per server so you're talking about 1.2 to 5 million USD a month. Perhaps you can make a better deal with that many servers ofcourse.

 

No need to say that's quite an enterprise job. If the viewers aren't simultaneous but just 2 million visitors per month, each viewing one video then you can

count how many servers you need with bitrate and length of the video.

 

E.g. 800 ppl can watch simultaneous on your server. Your videos have an average length of 10 minutes (youtube concept). You can serve 800 users * ( 60 minutes per hour / 10 minutes length) = 4800 videos per hour. If you visitors would be spread out over 10 hours of a day (instead of 24 hours a day) you could serve 4800 * 10 = 48.000 videos per day. Let's reserve some days and say 25 active days per month and your server would do 48.000 * 25 = 1.2 million video-views each month.

 

You could increase the maximum amount of simultaneous viewers by adding multiple 1 gbit nics in one server but even a fast quad core CPU will be your bottleneck.

 

My 2 cents,

 

Walter

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

---

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

---

Internal Virus Database is out-of-date.
Checked by AVG.
Version: 7.5.560 / Virus Database: 270.12.26/2116 - Release Date: 15-05-09 06:16

---

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

[Bobby-46]

I agree.  Red5 should explicitly take into account the Content Distribution Network of Rackspace.

There is an API to interface with the Cloud Files, so I'm not sure why it would be a problem to create a container for each new client you have (automatically too; by registration, payment, then API creation of container; then just put their videos in the container and they will be distributed on the Cloud CDN, put in the Terms of Service for the users that they must pay for the Bandwidth plus DISTRIBUTION & HANDLING --- Sound familiar?  Microeconomics...that goof ball selling PC learning software,  I'll give it to you FREE, you just pay SHIPPING & HANDLING :-) 

On Jul 2, 2009, at 8:27 PM, Andy Shaules wrote:

Live stream: power horse server with heap-big-pipe.

average 600 kbs  @ 350 viewers maximum.

 

My recommended high-quality, long-term, high-demand, no-crashes, 'Ican go to sleep' reliability setting.

 

More viewers than that, I burn out my drive belt because high gear is taller than the default settings can handle.(rtmp cuhunk size, and thread pool limits).

 

Can it handle more? yes...... Would I personally sell a single red5 server instance at a higher capacity capability which is used for HIGH quality streams? No... I would sugest distibuted deployment. 

 

:)

 

----- Original Message -----

From: walter@...

To: red5@...

Sent: Thursday, July 02, 2009 5:03 PM

Subject: Re: [Red5] short update on performance and Red5

Since I didn't test it I don't want to make wrong assumptions.

 

More tests coming up this weekend perhaps.

 

In theory Red5 should have less problems with live-streams since it only needs to buffer a few packets in-memory in contrast to VODs which need to be read from disc.

 

W.

 

Hey Walter, thanks for sharing this analysis. Do you think this only applies to to VOD and not to live  streams for any particular reason? Shouldn't we expect a similar result?

Txs

On Thu, Jul 2, 2009 at 8:35 PM, Walter Tak <walter@...> wrote:

Changes in Italic since original message was a bit sloppy.

 

For those that are a little interested in the performance of Red5 and VOD (video on demand e.g. no live-streams) then here are some measured timings:

 

I executed a few times 10 threads that each downloads as fast as possible 1 single VOD file ( 60 Mb ) via RTMP.

The downloading client is a dedicated quad-core xeon @ 2.13 ghz (X3210) on a 100 mbit link , hosted in a datacentre.

 

The Red5-server serving the files has the same configuration and had a load of 17% CPU (or 75% on one CPU) during the process while outputting at maximum network-capable speed (e.g. 100 mbit , 11 Mbytes). Before and after the process the cpu dropped to a around 0%.

 

10 threads each downloading a different VOD file (each between 30-60 Mb each) didn't differ at all ; same speed (max network) and same load.

 

Based on that very short test one could say that the same configuration , with a 1 gbit NIC instead of 100 mbit, could serve around 5 times that output (e.g. 500 mbit) before maxing out with CPU.

 

With a faster CPU (newer Xeons) that run around 3 Ghz and faster FSB's that value might even be increased with 50 to 100%. You might hit 750-900 Mbit per server,

in theory. 100 mbit continous streaming with a few streams didn't make my server sweat for sure.

 

So if you want to calculate "how many ppl can watch my videos before my servers are maxxed out" then just divide bandwidth by bitrate. If your box is powerful enough (simple dual core should be able to handle 100 mbit without problems) and you have the 100 mbit really dedicated uplink then you could serve 100 streams of 1 mbit each, or 1000 streams of 100 kbit each. (Or a mix ofcourse).

 

With a real 1 gbit connection (make sure you have someone paying the bills for that) you could increase that per box with a factor 5-8.

 

With a video of 1 mbit bitrate you can show decent H.264 content,near fullscreen, with a faster server and a 1 gbit NIC , you could do around 500-800 per box. Let's assume 800 VODs, each 1mbit, fast quad-xeon with a 1gbit NIC.

 

Someone asked for 2 million viewers (right?) ; if those were simultaneous you'd need 2.000.000 / 800 = 2500 servers. Each pumping out a massive 800-900 Mbit. Pricing for 1gbit flatfee bandwidth could approach 500-2000 usd per server so you're talking about 1.2 to 5 million USD a month. Perhaps you can make a better deal with that many servers ofcourse.

 

No need to say that's quite an enterprise job. If the viewers aren't simultaneous but just 2 million visitors per month, each viewing one video then you can

count how many servers you need with bitrate and length of the video.

 

E.g. 800 ppl can watch simultaneous on your server. Your videos have an average length of 10 minutes (youtube concept). You can serve 800 users * ( 60 minutes per hour / 10 minutes length) = 4800 videos per hour. If you visitors would be spread out over 10 hours of a day (instead of 24 hours a day) you could serve 4800 * 10 = 48.000 videos per day. Let's reserve some days and say 25 active days per month and your server would do 48.000 * 25 = 1.2 million video-views each month.

 

You could increase the maximum amount of simultaneous viewers by adding multiple 1 gbit nics in one server but even a fast quad core CPU will be your bottleneck.

 

My 2 cents,

 

Walter

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

---

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

---

Internal Virus Database is out-of-date.
Checked by AVG. 
Version: 7.5.560 / Virus Database: 270.12.26/2116 - Release Date: 15-05-09 06:16

---

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org

_______________________________________________
Red5 mailing list
Red5@...
http://osflash.org/mailman/listinfo/red5_osflash.org