Networking optimize

Meet VoIP bandwidth requirements without crippling your network performance

If you don't consider your bandwith requirements when building a VoIP infrastructure, you'll wind up with an unreliable system and an overtaxed WAN. George Ou looks at audio compression algorithms and WAN topologies to help you effectively determine your bandwidth needs.

This article is also available as a download that includes an Excel template you can use to run your own calculations.

VoIP bandwidth consumption over a WAN (wide area network) is one of the most important factors to consider When building a VoIP infrastructure.  Failure to account for VoIP bandwidth requirements will severely limit the reliability of a VoIP system and place a huge burden on the WAN infrastructure.  This article will teach you how to deal with the various audio compression algorithms and WAN topologies.  Once you learn how to control and calculate bandwidth utilization, you can set the proper amount of bandwidth allocation for QoS.

 

VoIP codec types

We'll start with the most common VoIP codec (compression/decompression) algorithms and examine their bandwidth and quality characteristics.  Note that MOS stands for mean opinion score, which is a subjective rating of sound quality.

Narrowband VoIP codecs
Codec Payload bandwidth MOS Description
G.711 64 kbps 4.1 This is the most universally supported codec used in IP telephony.  This narrowband codec supports frequencies in the 300 to 3,400 hertz range and is uncompressed. Although the quality is very good, it consumes a lot of bandwidth.
G.729 8 kbps 3.92 This is the second most supported codec and offers nearly the same quality as G.711. The key advantage is that it is compressed eight times smaller than G.711 while sounding almost as good.

Wideband VoIP codecs
Codec Payload bandwidth Description
G.722 48 to 64 kbps This is the most common wideband codec available in IP phones, though wideband support is only recently gaining momentum. The quality is excellent at twice the sampling rate of standard G.711, but the compression isn't that great. But considering the fact that it's the same bit rate as narrowband G.711 but delivers much more realistic sound, G.722 will be one of the best codecs to use if your IP telephones will support it. Wideband supports frequencies of 50 to 7,000 hertz. There are no longer any patents covering G.722, so it's free for anyone to use.
G.722.1 16 to 32 kbps This is a wideband codec, aka Siren7, developed by Polycom. Its key advantage is that it's a computationally efficient and compact codec at 16 to 32 kbps, which is less than half the bandwidth required by G.722.  16 kbps mode isn't appropriate for noisy audio input or if music is mixed in, since the compression artifacts are noticeable.  32 kbps is good for any kind of workload.  This codec must be licensed from Polycom, and it's currently used only in Polycom's high-end video conferencing systems under the marketing name "Ultimate HD." Current Polycom IP phones use the marketing term "HD Voice," which supports only generic G.722 for its wideband codec, although future IP phone models may support G.722.1.
G.722.2 6.6 to 23.85 kbps Also known as AMR-WB, this is a wideband codec. A 6.6 kbps mode is also supported, but 12.75 is the practical bit rate for speech in a clean environment. The higher 23.85 bit rate is better for noisy conditions and music. At the time of this writing, I'm not aware of any IP phones that support this codec. It's currently used by T-Mobile in Germany for cell phone applications.
Speex wideband 10 to 28 kbps This is an excellent open source codec from Speex that offers very good wideband quality at relatively low bit rates. VBR (variable bit rate) is also supported in 12 or 18 kbps mode. This is a free open source codec and can be used by anyone. It's supported by open source Asterisk PBX, but no hard IP phones and only one soft IP phone support it.
Siren14 ultra-wideband 24 to 32 kbps Siren14 is a free-to-license (not to be confused with license-free) ultra-wideband version of the G.722.1 codec from Polycom. Some of the high-end features in G.722.1, like echo cancellation and noise reduction, are omitted. Even though this is royalty free, you still need a license from Polycom. Siren14 supports a wider dynamic range up to 14,000 hertz, compared with 7,000 hertz in wideband.
Siren22 ultra-wideband 32 to 64 kbps Siren22 is a proprietary ultra-wideband codec from Polycom that currently can't be licensed. Siren22 supports an even wider dynamic range--up to 22,000 hertz. The digital sampling rate is 48 KHz.

The bottom line is that G.711 and G.729 will be the lowest common denominators that are universally supported. G.729 makes the most sense in the narrowband range because it sounds almost as good as uncompressed G.711 but is eight times smaller. Since inexpensive G.722 phones are now on the market in the $80 to $400 range, anyone willing to expend 64 kbps should avoid G.711 and use wideband G.722 as much as possible because no extra bandwidth is required.  I have seen too many VoIP implementations that blindly use G.711 without any consideration to bandwidth.

There are plenty of nice, free wideband codecs that offer excellent sound quality at half the bit rate of G.722, but getting some, let alone all, of the IP phone manufacturers on board to use them is difficult. It's hard enough just getting the phone makers to support generic G.722. Let's hope that they all come around in the next few years and start supporting all the free wideband codecs.

 


Free network administration newsletter

Need help configuring, administering, supporting, and optimizing network infrastructure? Turn to our free Network Administration NetNote. Automatically sign up today!


 

The packet overhead tax

It isn't enough to look at just the raw payload bandwidth used by the various codecs. We must also consider the packet overhead from the various types of connections. Since VoIP never consumes more than 90 kbps on an Ethernet line, there are almost never any bandwidth problems on a LAN (local area network). Even a 10 mbps switched network can carry 100 simultaneous channels of the heaviest codecs. The problem arrives when VoIP packets must traverse the WAN or Internet, where bandwidth is typically limited from 64 kbps to 1,536 kbps in a T1 connection. Although many DSL or cable broadband connections boast 6 mbps or greater throughput, their upstream speed is the big limiting factor, at 128 kbps to 1,000 kbps. For connections that are limited to 64 kbps (often the case for retail chains) and that must also support data applications, special measures must be taken to make VoIP possible.

Packet overhead can be a severe problem if left unchecked. This is because at least 50 VoIP packets must be sent per second to keep the packetization delay to a minimum. Since the IP, UDP, and RTP headers in each VoIP packet are 40 bytes (320 bits), 50 of them in one second means it will consume 16 kbps. The physical transport mechanism and the protocols they use will bump up the overhead even more.

On a frame relay or T1 connection, the typical packet overhead is 18.8 kbps per channel. This may not be a big issue for 64 kbps codecs like G.711 or G.722, since it's relatively small compared to the payload. But for 8 to 32 kbps codecs, like G.729 and G.722.1, 18.8 kbps is huge by comparison. Packet overhead for site-to-site IPSEC VPN connections is far more severe. The overhead can go from 35 kbps to 45 kbps per channel depending on the type of VPN implemented. That means every single call over a VPN tunnel gets a ~40 kbps tax slapped on top of the codec bit rate.

 

Dealing with packet overhead

There are two main ways of dealing with packet overhead, depending on the network topology. We can try to reduce the size of the packet headers with compression or we can trunk multiple voice streams into a single packet. Packet header compression can reduce the IP/UDP/RDP header by a factor of 10 to 20 times, but it works only with frame relay or MP (multilink point-to-point protocol) connections, and both sides of the connection must be configured for it. Packet header compression is the ideal solution whenever it can be employed.

MPLS connections can't use header compression because they aren't point-to-point connections and must go over a large virtual cloud through multiple carriers. MPLS does have the advantage of a mesh topology, where any site can talk to any other site directly without relaying through some central hub site. This can reduce latency in some situations. VoIP calls over the Internet also can't use packet header compression. Site-to-site IPSEC VPN connections not only can't use packet header compression, they exacerbate the situation by doubling the overhead. The only thing we can do for MPLS and VPN WAN topology is use VoIP packet trunking.

VoIP trunking is a method of taking multiple VoIP streams and merging them into a single stream sharing a single packet header. So instead of having to multiply the packet overhead bit rate by the number of channels, it's a one-time fixed cost for any number of channels with the same source and destination. This can be done only using an open source Asterisk PBX gateway in each remote site.

This method works even if you're using a proprietary VoIP PBX and IP phones, so long as they can speak the SIP or H.323 protocols. Asterisk can translate them to its own efficient IAX2 protocol, which supports channel trunking, and retranslate to SIP or H.323 once it emerges from the limited bandwidth connection. VoIP trunking is the ideal solution whenever header compression can't be used. In future articles, we'll look at how to implement these solutions.

 

Case studies and examples

Let's consider an example in which the WAN topology consists mostly of 64 to 256 kbps frame relay or MP connections. In this scenario, we must primarily use narrowband G.729 8 kbps codecs, and header compression is essential. We can support a single high-end Polycom HD Voice conference phone using the G.722.1 Siren7 codec at 32 kbps, which is justifiable when an important meeting is taking place. Here is a breakdown of how much bandwidth for a given number of channels is supported.

Frame relay or MP WAN topology with header compression
Channels Overhead Codec bit rate Trunked bit rate Unicast bit rate
4 3.6 kbps G.729 (8 kbps) 35.6 kbps 46.4 kbps
8 3.6 kbps G.729 (8 kbps) 67.6 kbps 92.8 kbps
16 3.6 kbps G.729 (8 kbps) 131.6 kbps 185.6 kbps
1 3.6 kbps G.722.1 (32 kbps) 35.6 kbps 35.6 kbps

Although the ROI on trunking doesn't seem like much to justify building an Asterisk gateway, we can use that same Linux appliance as a transparent proxy and HTTP content compressor. This allows us to get a lot more out of our WAN connection. We can sometimes reduce the data bandwidth consumption by a factor of 4 or 10.

Now, let's see what happens in an MPLS WAN or Internet topology.

MPLS WAN or Internet topology
Channels Overhead Codec bit rate Trunked Unicast
8 18.8 kbps G.729 (8 kbps) 82.8 kbps 214.4 kbps
8 18.8 kbps G.711 (64 kbps) 530.8 kbps 662.4 kbps
8 18.8 kbps G.722.1 (32 kbps) 274.8 kbps 406.4 kbps
8 18.8 kbps G.722 (64 kbps) 530.8 kbps 662.4 kbps

As you can see, VoIP trunking is the only way to control bandwidth usage for the 8 and 32 kbps codecs, and an Asterisk PBX gateway is the only way to solve this problem.

Finally, lets see what happens in an IPSEC VPN site-to-site WAN topology. Note that we're using only an IPSEC tunnel and we're not using GRE on top, so that we don't increase the packet overhead any further.

IPSEC VPN site-to-site WAN topology
Channels Overhead Codec bit rate Trunked Unicast
8 34.8 kbps G.729 (8 kbps) 98.8 kbps 342.4 kbps
8 34.8 kbps G.711 (64 kbps) 546.8 kbps 790.4 kbps
8 34.8 kbps G.722.1 (32 kbps) 290.8 kbps 534.4 kbps
8 34.8 kbps G.722 (64 kbps) 546.8 kbps 790.4 kbps

IPSEC packet overhead is the most severe of all the topologies, but it can be brought under control with VoIP trunking. IPSEC connections are the cheapest of all WAN topologies because inexpensive high-speed broadband connections (less than $100 per month) can be used. Relatively cheap ISDN (MP 128 kbps) or MPLS connections can be used to provide reasonably cheap redundancy, but none of this will be practical if we ignore the packet overhead tax and use the wrong VoIP codecs.

If you want to run your own calculations, the download version of this article includes an Excel template you can use.

7 comments
mike82
mike82

What bandwidth does a phone use simply registering with the server. Or when i pick up the phone and get a dialtone but not place a call. Is it always using the codec bandwidth even if idle?

anandaramesh
anandaramesh

Hi, this is really an informative one and helpful to calculate the bandwidth required. I have couple of questions here. 1. How did you calculated the Bandwidth overhead? 2. Is these calculations are for oneway communication or two way?

peter
peter

Very informative article. Thankyou for taking the time to put this together.

guillermo_mansilla
guillermo_mansilla

Hi, I read your article, it is very interesting but I didn't learn how much bandwidth is enough and how many video conference calls can be made given a bandwidth. Could you answer that? Thank very much

JodyGilbert
JodyGilbert

Have you run into problems setting the proper bandwidth allocation for QoS? Are there any aspects of determining your bandwidth needs that you find confusing?

georgeou
georgeou

Packet headers are always the same size regardless of the payload. It's easy to count the packets per second and the header sizes. Yes, bandwidth consumption is two-way.

georgeou
georgeou

The bandwidth requirements vary greatly depending on the network topology and the codec you use. That is what the charts on the bottom explain.