Hello,
Jan Janak conducted a very interesting research project regarding energy efficiency of VoIP systems during 2010, a collaboration between iptel.org and Columbia University.
The team used the source code from sip-router.org GIT repository from January 2010, which corresponds to Kamailio (former OpenSER) and SER v3.0. The latest stable series v3.1 shares the same internal architecture with v3.0.
As part of the research work, Jan could also gather some figures about capacity and performances of v3.0 with a quite complex configuration file: etc/sip-router-oob.cfg (involving authentication and NAT traversal as well).
You can read the paper about energy efficiency at:
- Green VoIP Article: http://asipto.com/u/2j
The draft notes about capacity and performances of v3.0 are available at:
- Performances and Capacity for v3.0 Wiki page: http://asipto.com/u/2k
Some interesting results:
- one instance of SIP server with 500 000 online users (mixed users – behind and not NAT routers) – consumed energy 210W - one instance of SIP server with 1 000 000 online users (no NAT involved) – consumed energy 190W - on a 32-bit machine with 4GB of memory and with 2.5GB reserved for SIP server, the server could support 43 000 simultaneous TLS connections – consumed energy 203W - one SIP server instance with 80 000 permanent TCP connections, the SIP server could still handle at least 1000 requests per second and a connection arrival rate of 1000 new connections per second, done for 20 000 new connections. CPU load generated by the SIP server was from 6% to 8%.
I added a new section to the draft notes to list the enhancements done for the latest stable release (v3.1.x) that contribute to performance improvements, like asynchronous TLS, fine tuning of memory for TLS connections and raw UDP sockets.
Cheers, Daniel
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL & switches is where the power savings are important.
Daniel-Constantin Mierla wrote:
Hello,
Jan Janak conducted a very interesting research project regarding energy efficiency of VoIP systems during 2010, a collaboration between iptel.org and Columbia University.
The team used the source code from sip-router.org GIT repository from January 2010, which corresponds to Kamailio (former OpenSER) and SER v3.0. The latest stable series v3.1 shares the same internal architecture with v3.0.
As part of the research work, Jan could also gather some figures about capacity and performances of v3.0 with a quite complex configuration file: etc/sip-router-oob.cfg (involving authentication and NAT traversal as well).
You can read the paper about energy efficiency at:
- Green VoIP Article: http://asipto.com/u/2j
The draft notes about capacity and performances of v3.0 are available at:
- Performances and Capacity for v3.0 Wiki page: http://asipto.com/u/2k
Some interesting results:
- one instance of SIP server with 500 000 online users (mixed users –
behind and not NAT routers) – consumed energy 210W
- one instance of SIP server with 1 000 000 online users (no NAT
involved) – consumed energy 190W
- on a 32-bit machine with 4GB of memory and with 2.5GB reserved for
SIP server, the server could support 43 000 simultaneous TLS connections – consumed energy 203W
- one SIP server instance with 80 000 permanent TCP connections, the
SIP server could still handle at least 1000 requests per second and a connection arrival rate of 1000 new connections per second, done for 20 000 new connections. CPU load generated by the SIP server was from 6% to 8%.
I added a new section to the draft notes to list the enhancements done for the latest stable release (v3.1.x) that contribute to performance improvements, like asynchronous TLS, fine tuning of memory for TLS connections and raw UDP sockets.
Cheers, Daniel
On Wednesday 25 May 2011, Jeremya wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL & switches is where the power savings are important.
Hi Jeremya,
you're right in this case - given the much higher number of CPE devices in the end user hands and especially as most of them use really cheap (also bad) power supplies. But in generally optimisations on a server level are also important, if you have a few thousand servers than even small amounts fast add up. And for the whole internet it already uses siginifcant part of the power, close to 10% in the US to some research:
http://hardware.slashdot.org/story/07/09/27/2157230/Internet-Uses-94-of- Electricity-In-the-US
With regards to the research from Jan, the performance numbers are really impressive, great work!
Cheers,
Henning
Hello Henning,
My application is slightly different in that I have hundreds of nodes each running an instance of Kamailio on embedded servers in a failure tolerant configuration. This is in a risk-of-life application and each node has large backup power supplies.
In reality kamailio takes milliwatts of the total load. What sucks power at each node is the SIP devices, switches, and the CISCO routers foisted on me by the client. As an engineer (? I guess I am?) The boring stuff becomes important so total power draw and life after power failure is very important. Without the power hungry devices my kamailio systems can stay up for weeks after power loss. With the SIP devices and routers I'm limited to hours of uptime.
Henning Westerholt wrote:
On Wednesday 25 May 2011, Jeremya wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL & switches is where the power savings are important.
Hi Jeremya,
you're right in this case - given the much higher number of CPE devices in the end user hands and especially as most of them use really cheap (also bad) power supplies. But in generally optimisations on a server level are also important, if you have a few thousand servers than even small amounts fast add up. And for the whole internet it already uses siginifcant part of the power, close to 10% in the US to some research:
http://hardware.slashdot.org/story/07/09/27/2157230/Internet-Uses-94-of- Electricity-In-the-US
With regards to the research from Jan, the performance numbers are really impressive, great work!
Cheers,
Henning
I should point out that my SIP business is risk-of-life voice services where I have to arrange to power SIP devices and communications devices from alternate energy sources for extended periods, so I am intimately aware of every watt that is used in SIP communications.
My systems use kamailio in embedded servers at very low power. My constraint is the SIP devices and communications devices that suck up an order of magnitude more power at each small site - of which I have many.
Jeremya wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL & switches is where the power savings are important.
Daniel-Constantin Mierla wrote:
Hello,
Jan Janak conducted a very interesting research project regarding energy efficiency of VoIP systems during 2010, a collaboration between iptel.org and Columbia University.
The team used the source code from sip-router.org GIT repository from January 2010, which corresponds to Kamailio (former OpenSER) and SER v3.0. The latest stable series v3.1 shares the same internal architecture with v3.0.
As part of the research work, Jan could also gather some figures about capacity and performances of v3.0 with a quite complex configuration file: etc/sip-router-oob.cfg (involving authentication and NAT traversal as well).
You can read the paper about energy efficiency at:
- Green VoIP Article: http://asipto.com/u/2j
The draft notes about capacity and performances of v3.0 are available at:
- Performances and Capacity for v3.0 Wiki page: http://asipto.com/u/2k
Some interesting results:
- one instance of SIP server with 500 000 online users (mixed users –
behind and not NAT routers) – consumed energy 210W
- one instance of SIP server with 1 000 000 online users (no NAT
involved) – consumed energy 190W
- on a 32-bit machine with 4GB of memory and with 2.5GB reserved for
SIP server, the server could support 43 000 simultaneous TLS connections – consumed energy 203W
- one SIP server instance with 80 000 permanent TCP connections, the
SIP server could still handle at least 1000 requests per second and a connection arrival rate of 1000 new connections per second, done for 20 000 new connections. CPU load generated by the SIP server was from 6% to 8%.
I added a new section to the draft notes to list the enhancements done for the latest stable release (v3.1.x) that contribute to performance improvements, like asynchronous TLS, fine tuning of memory for TLS connections and raw UDP sockets.
Cheers, Daniel
SIP Express Router (SER) and Kamailio (OpenSER) - sr-users mailing list sr-users@lists.sip-router.org http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-users
On Wed, May 25, 2011 at 06:54, Jeremya jeremy@electrosilk.net wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL & switches is where the power savings are important.
Sure, the total power consumption of the whole system is dominated by the power consumption of end-point devices, there's no doubt about that and the paper says that.
Nevertheless, as an ITSP you are typically paying for the energy consumed by your servers and in that case knowing what you can expect and how many servers you need is useful. Modern data-center servers have significant base-line power consumption and a portion of that needs to be attributed to the SIP service running on those servers.
-Jan
Daniel-Constantin Mierla wrote:
Hello,
Jan Janak conducted a very interesting research project regarding energy efficiency of VoIP systems during 2010, a collaboration between iptel.org and Columbia University.
The team used the source code from sip-router.org GIT repository from January 2010, which corresponds to Kamailio (former OpenSER) and SER v3.0. The latest stable series v3.1 shares the same internal architecture with v3.0.
As part of the research work, Jan could also gather some figures about capacity and performances of v3.0 with a quite complex configuration file: etc/sip-router-oob.cfg (involving authentication and NAT traversal as well).
You can read the paper about energy efficiency at:
- Green VoIP Article: http://asipto.com/u/2j
The draft notes about capacity and performances of v3.0 are available at:
- Performances and Capacity for v3.0 Wiki page: http://asipto.com/u/2k
Some interesting results:
- one instance of SIP server with 500 000 online users (mixed users –
behind and not NAT routers) – consumed energy 210W
- one instance of SIP server with 1 000 000 online users (no NAT
involved) – consumed energy 190W
- on a 32-bit machine with 4GB of memory and with 2.5GB reserved for
SIP server, the server could support 43 000 simultaneous TLS connections – consumed energy 203W
- one SIP server instance with 80 000 permanent TCP connections, the
SIP server could still handle at least 1000 requests per second and a connection arrival rate of 1000 new connections per second, done for 20 000 new connections. CPU load generated by the SIP server was from 6% to 8%.
I added a new section to the draft notes to list the enhancements done for the latest stable release (v3.1.x) that contribute to performance improvements, like asynchronous TLS, fine tuning of memory for TLS connections and raw UDP sockets.
Cheers, Daniel
SIP Express Router (SER) and Kamailio (OpenSER) - sr-users mailing list sr-users@lists.sip-router.org http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-users
Hello,
out of curiosity, since you used the sources from GIT - was memory debugging on? It is usually enabled in master branch and that could have some impact in memory usage and performances...
Thanks, Daniel
On 5/25/11 3:00 PM, Jan Janak wrote:
On Wed, May 25, 2011 at 06:54, Jeremyajeremy@electrosilk.net wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL& switches is where the power savings are important.
Sure, the total power consumption of the whole system is dominated by the power consumption of end-point devices, there's no doubt about that and the paper says that.
Nevertheless, as an ITSP you are typically paying for the energy consumed by your servers and in that case knowing what you can expect and how many servers you need is useful. Modern data-center servers have significant base-line power consumption and a portion of that needs to be attributed to the SIP service running on those servers.
-Jan
Daniel-Constantin Mierla wrote:
Hello,
Jan Janak conducted a very interesting research project regarding energy efficiency of VoIP systems during 2010, a collaboration between iptel.org and Columbia University.
The team used the source code from sip-router.org GIT repository from January 2010, which corresponds to Kamailio (former OpenSER) and SER v3.0. The latest stable series v3.1 shares the same internal architecture with v3.0.
As part of the research work, Jan could also gather some figures about capacity and performances of v3.0 with a quite complex configuration file: etc/sip-router-oob.cfg (involving authentication and NAT traversal as well).
You can read the paper about energy efficiency at:
- Green VoIP Article: http://asipto.com/u/2j
The draft notes about capacity and performances of v3.0 are available at:
- Performances and Capacity for v3.0 Wiki page: http://asipto.com/u/2k
Some interesting results:
- one instance of SIP server with 500 000 online users (mixed users –
behind and not NAT routers) – consumed energy 210W
- one instance of SIP server with 1 000 000 online users (no NAT
involved) – consumed energy 190W
- on a 32-bit machine with 4GB of memory and with 2.5GB reserved for
SIP server, the server could support 43 000 simultaneous TLS connections – consumed energy 203W
- one SIP server instance with 80 000 permanent TCP connections, the
SIP server could still handle at least 1000 requests per second and a connection arrival rate of 1000 new connections per second, done for 20 000 new connections. CPU load generated by the SIP server was from 6% to 8%.
I added a new section to the draft notes to list the enhancements done for the latest stable release (v3.1.x) that contribute to performance improvements, like asynchronous TLS, fine tuning of memory for TLS connections and raw UDP sockets.
Cheers, Daniel
SIP Express Router (SER) and Kamailio (OpenSER) - sr-users mailing list sr-users@lists.sip-router.org http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-users
SIP Express Router (SER) and Kamailio (OpenSER) - sr-users mailing list sr-users@lists.sip-router.org http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-users
No, I turned it off.
-Jan
On Thu, May 26, 2011 at 11:50, Daniel-Constantin Mierla miconda@gmail.com wrote:
Hello,
out of curiosity, since you used the sources from GIT - was memory debugging on? It is usually enabled in master branch and that could have some impact in memory usage and performances...
Thanks, Daniel
On 5/25/11 3:00 PM, Jan Janak wrote:
On Wed, May 25, 2011 at 06:54, Jeremyajeremy@electrosilk.net wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL& switches is where the power savings are important.
Sure, the total power consumption of the whole system is dominated by the power consumption of end-point devices, there's no doubt about that and the paper says that.
Nevertheless, as an ITSP you are typically paying for the energy consumed by your servers and in that case knowing what you can expect and how many servers you need is useful. Modern data-center servers have significant base-line power consumption and a portion of that needs to be attributed to the SIP service running on those servers.
-Jan
Daniel-Constantin Mierla wrote:
Hello,
Jan Janak conducted a very interesting research project regarding energy efficiency of VoIP systems during 2010, a collaboration between iptel.org and Columbia University.
The team used the source code from sip-router.org GIT repository from January 2010, which corresponds to Kamailio (former OpenSER) and SER v3.0. The latest stable series v3.1 shares the same internal architecture with v3.0.
As part of the research work, Jan could also gather some figures about capacity and performances of v3.0 with a quite complex configuration file: etc/sip-router-oob.cfg (involving authentication and NAT traversal as well).
You can read the paper about energy efficiency at:
- Green VoIP Article: http://asipto.com/u/2j
The draft notes about capacity and performances of v3.0 are available at:
- Performances and Capacity for v3.0 Wiki page: http://asipto.com/u/2k
Some interesting results:
- one instance of SIP server with 500 000 online users (mixed users –
behind and not NAT routers) – consumed energy 210W
- one instance of SIP server with 1 000 000 online users (no NAT
involved) – consumed energy 190W
- on a 32-bit machine with 4GB of memory and with 2.5GB reserved for
SIP server, the server could support 43 000 simultaneous TLS connections – consumed energy 203W
- one SIP server instance with 80 000 permanent TCP connections, the
SIP server could still handle at least 1000 requests per second and a connection arrival rate of 1000 new connections per second, done for 20 000 new connections. CPU load generated by the SIP server was from 6% to 8%.
I added a new section to the draft notes to list the enhancements done for the latest stable release (v3.1.x) that contribute to performance improvements, like asynchronous TLS, fine tuning of memory for TLS connections and raw UDP sockets.
Cheers, Daniel
SIP Express Router (SER) and Kamailio (OpenSER) - sr-users mailing list sr-users@lists.sip-router.org http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-users
SIP Express Router (SER) and Kamailio (OpenSER) - sr-users mailing list sr-users@lists.sip-router.org http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-users
-- Daniel-Constantin Mierla http://www.asipto.com
Jan-
On Wed, May 25, 2011 at 06:54, Jeremya jeremy@electrosilk.net wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL & switches is where the power savings are important.
Sure, the total power consumption of the whole system is dominated by the power consumption of end-point devices, there's no doubt about that and the paper says that.
Nevertheless, as an ITSP you are typically paying for the energy consumed by your servers and in that case knowing what you can expect and how many servers you need is useful. Modern data-center servers have significant base-line power consumption and a portion of that needs to be attributed to the SIP service running on those servers.
Just want to clarify... I assume that no transcoding, transrating, or other "translation" between end points is included? I don't see any mention of rtpproxy or other media servers. I ask because these tend to be compute-intensive tasks that would have significant impact server energy usage and performance (e.g. max calls and calls-per-sec).
-Jeff
Daniel-Constantin Mierla wrote:
Hello,
Jan Janak conducted a very interesting research project regarding energy efficiency of VoIP systems during 2010, a collaboration between iptel.org and Columbia University.
The team used the source code from sip-router.org GIT repository from January 2010, which corresponds to Kamailio (former OpenSER) and SER v3.0. The latest stable series v3.1 shares the same internal architecture with v3.0.
As part of the research work, Jan could also gather some figures about capacity and performances of v3.0 with a quite complex configuration file: etc/sip-router-oob.cfg (involving authentication and NAT traversal as well).
You can read the paper about energy efficiency at:
- Green VoIP Article: http://asipto.com/u/2j
The draft notes about capacity and performances of v3.0 are available at:
- Performances and Capacity for v3.0 Wiki page: http://asipto.com/u/2k
Some interesting results:
- one instance of SIP server with 500 000 online users (mixed users â
behind and not NAT routers) â consumed energy 210W
- one instance of SIP server with 1 000 000 online users (no NAT
involved) â consumed energy 190W
- on a 32-bit machine with 4GB of memory and with 2.5GB reserved for
SIP server, the server could support 43 000 simultaneous TLS connections â consumed energy 203W
- one SIP server instance with 80 000 permanent TCP connections, the
SIP server could still handle at least 1000 requests per second and a connection arrival rate of 1000 new connections per second, done for 20 000 new connections. CPU load generated by the SIP server was from 6% to 8%.
I added a new section to the draft notes to list the enhancements done for the latest stable release (v3.1.x) that contribute to performance improvements, like asynchronous TLS, fine tuning of memory for TLS connections and raw UDP sockets.
Cheers, Daniel
On Fri, May 27, 2011 at 23:43, Jeff Brower jbrower@signalogic.com wrote:
Jan-
On Wed, May 25, 2011 at 06:54, Jeremya jeremy@electrosilk.net wrote:
These figures pale into insignificance compared to the power required for standard SIP devices - typically 5-8 watts per device multiplied by the number of devices.
When you factor in Gigabit Ethernet the power ups significantly.
Optimisation at the server level is not significant on any scale. Optimisation on communications power: i.e. end-devices, DSL & switches is where the power savings are important.
Sure, the total power consumption of the whole system is dominated by the power consumption of end-point devices, there's no doubt about that and the paper says that.
Nevertheless, as an ITSP you are typically paying for the energy consumed by your servers and in that case knowing what you can expect and how many servers you need is useful. Modern data-center servers have significant base-line power consumption and a portion of that needs to be attributed to the SIP service running on those servers.
Just want to clarify... I assume that no transcoding, transrating, or other "translation" between end points is included? I don't see any mention of rtpproxy or other media servers. I ask because these tend to be compute-intensive tasks that would have significant impact server energy usage and performance (e.g. max calls and calls-per-sec).
Correct, those tests did not include any manipulation of media streams. We were primarily interested in system components that are present in SIP based deployments and missing in Skype because we wanted to see how those two systems differ in terms of architecture, power consumption, etc.
We did run some tests with rtpproxy-like media relays though. You'll find some numbers in the paper.
-Jan
-
Daniel-Constantin Mierla -
Henning Westerholt -
Jan Janak -
Jeff Brower -
Jeremya