From the network's point of view, a live broadcast is traffic that cannot wait. A file download tolerates a packet that is delayed or has to be resent; at worst the viewer gets the last byte a few seconds later. A video picture is different. With a 25 fps signal it expects a new frame every 40 milliseconds, and if the frame does not arrive in time, the decoder shows a blocky, frozen or black picture.
Event venue networks are also rarely built for this. Ticket scanning, the press Wi-Fi, the catering card terminals and the camera feeds run on the same switch, and usually nobody has measured the upload bandwidth with the house full.
How much bandwidth a broadcast needs
The answer differs by orders of magnitude depending on the form the signal travels in. In studios and OB vans, video has traditionally run over SDI cable: 3G-SDI (SMPTE ST 424) runs at roughly 2.97 Gbit/s, and 12G-SDI (SMPTE ST 2082), used for 4K, at roughly 11.88 Gbit/s. If the same signal goes onto an IP network uncompressed, for example under the SMPTE ST 2110 family of standards, the network has to carry traffic of the same order of magnitude, because ST 2110-20 carries only the active picture and leaves out the blanking intervals. A simple calculation: 1920 × 1080 pixels, 20 bits per pixel (4:2:2 sampling, 10-bit color depth), 50 frames per second comes to roughly 2.1 Gbit/s for a single camera, and 4K is four times that.
At most events the picture travels on in compressed form. The encoder brings the bit rate down to a few or a few tens of Mbit/s using H.264 or H.265, and the target value is set in the encoder. You get exactly that amount of traffic only at constant bit rate (CBR). Between the two sits light compression with practically no visible loss, such as JPEG XS (ISO/IEC 21122, carried under ST 2110-22). It cuts the traffic to a fraction of the uncompressed signal while adding very little latency.
| Signal format | Order of magnitude of the bit rate | Typical use |
|---|---|---|
| Uncompressed (SDI, ST 2110-20) | Gbit/s per camera | Studio, OB van, on-site production |
| Lightly compressed (JPEG XS) | A few hundred Mbit/s | Production between sites with little latency headroom |
| Compressed (H.264, H.265) | A few or a few tens of Mbit/s | Uplink to the studio or to a streaming platform |
How to work it out
The bandwidth you need is the sum of the streams, plus packetization and error correction overhead. If, for example, four camera encoders are each set to 20 Mbit/s, that is 80 Mbit/s of payload. On top of that come a few percent for IP, UDP and RTP headers, the error correction packets or the bandwidth set aside for retransmission, the audio if it travels as a separate stream, and the traffic of the playback monitors or the director's line.
Do not compare the result with the nominal capacity of the line. A variable bit rate encoder briefly goes above the configured average during fast motion, such as a lighting effect or confetti. On many internet subscriptions the upload direction is much smaller than the download, and at the venue every other service shares the same line. Plan from the upload speed measured under load, and leave plenty of headroom.
Unicast and multicast, IGMP snooping
With unicast, the source sends a separate copy to each receiver. If five decoders in five rooms watch one encoder's signal, the encoder output carries five times the traffic. With few receivers this is not a problem, and internet uplinks work this way too: the encoder sends a single stream to the streaming platform, and the platform delivers it to the viewers.
With multicast (IETF RFC 1112), the source sends a single stream to a group address, and the network replicates it only where the path branches. The receiver uses IGMP (IGMPv2, RFC 2236, or IGMPv3, RFC 3376, on IPv4; MLDv2, RFC 3810, on IPv6) to signal that it wants a group's traffic. Distribution within the venue, video meant for many displays, and ST 2110 and AES67 media traffic are typically multicast.
Snooping and the querier
A simple switch treats multicast frames like broadcasts and sends them out of every port in the VLAN. A 2 Gbit/s stream then also lands on a 1 Gbit/s office port or a printer and clogs it. With IGMP snooping (RFC 4541), the switch listens in on the receivers' join messages and forwards a group's traffic only to the ports where someone has asked for it.
Snooping does not work reliably without a querier. The querier regularly asks who is still watching a given group, and the receivers' replies refresh the switches' tables. If nobody in the VLAN sends queries, the entries expire after a while, and a few minutes later the picture drops or, depending on the switch configuration, floods everywhere again. A designated querier in each VLAN is therefore as much a part of the design as snooping itself.
Other things to watch
- Address allocation. Only the lower 23 bits of an IPv4 group address are carried into the Ethernet MAC address, so 32 different group addresses map to the same MAC address. If the addresses of two streams collide, a device that filters only at the MAC level forwards both. Allocate group addresses deliberately, from a range set aside for the purpose.
- Routing. Multicast crosses a segment boundary only with multicast routing, typically PIM-SM (RFC 7761). Where the receiver knows which source it wants the stream from, source-specific multicast (SSM, RFC 4607) is simpler and more predictable.
- Wi-Fi. On a wireless network the access point sends multicast frames without acknowledgment and typically at a low data rate, so one large stream can take up the whole radio channel. Keep broadcast multicast off the Wi-Fi, and send the picture to wireless clients as unicast.
QoS: priority for broadcast traffic
QoS (Quality of Service) decides what happens when more packets are waiting at an output port than the port can carry. Packet marking and rate limiting work regardless of load, while queuing only matters once the port is full. When it does fill up, queuing decides whether the camera picture or a packet from a software update in progress gets dropped.
In the IP header, the DSCP field (RFC 2474) marks the packet's class. RFC 4594 recommends markings for common traffic types: CS3 for broadcast video, AF3x for non-interactive media streams, CS4 for real-time interactive video, and EF (value 46, RFC 3246) for voice. Many media networks use their own scheme instead. What matters is that the scheme is consistent and every switch interprets it the same way.
How it is built
- Marking at the source. The encoder, camera or audio device sets the DSCP value itself, or the access switch port does it for the device.
- Trust boundary. On broadcast ports the switch accepts the marking, and on every other port it overwrites it. Otherwise any laptop could put its own traffic in the highest class.
- Queues and scheduling. Switches and routers place packets in separate queues based on the marking. Give the strict priority queue an upper limit, because without one a large, wrongly marked stream can starve all other traffic.
- Verification. The queues' drop counters show whether the configuration needs changing anywhere.
QoS works inside your own network. Internet providers typically ignore DSCP marking, so on the uplink it helps if the venue limits its own outbound traffic and reserves room within it for the broadcast. QoS does not make up for capacity: if the broadcast traffic alone exceeds the line, priority only decides what gets lost.
Latency, jitter and packet loss
These are three separate things. Latency is the time a packet takes to get from the encoder to the decoder. Jitter is the variation in latency; the RTP standard (RFC 3550) measures it as the variation in the time differences between consecutively arriving packets. Packet loss is the share of packets that never arrive.
Jitter turns into latency
The decoder collects packets in a buffer and takes them out at a steady rate. The higher the jitter, the larger the buffer has to be, and the buffer size adds directly to the total latency. A packet that arrives later than the buffer allows is lost as far as the decoder is concerned. In a one-way broadcast, even a few seconds of latency bothers nobody. A live conversation, such as an interview patched in from a remote location, is another matter: ITU-T G.114 recommends at most 150 ms of one-way delay for interactive speech, and above that the two sides increasingly talk over each other.
Where jitter and loss come from
- Microbursts. Average utilization per minute is low, yet a port overflows for a few milliseconds and the switch's small buffer drops packets. A graph of the average does not show this; only the interface drop counter does.
- Speed changes. Where traffic steps down from a 10 Gbit/s link to a 1 Gbit/s port, it queues, and that is where latency varies the most.
- Wi-Fi in the middle of the signal path. On a radio channel, packets wait and get retransmitted, which makes latency vary. Connect cameras and encoders by cable.
- Faulty cable or port. CRC errors at one point in the network cause continuous low-level loss, which shows up in the broadcast as occasional picture errors.
What a lost packet does
In compressed video, most frames describe only what has changed since earlier frames. If the lost packet belongs to a frame that later frames refer to, the picture error can carry on until the next keyframe, and the viewer sees blocking or a frozen picture. There are two ways to protect against this. Forward error correction (FEC, for an MPEG-TS stream for example under SMPTE ST 2022-1) sends extra data up front, from which the receiver recovers small losses, at the cost of more traffic. Retransmission protocols, such as SRT or RIST as defined in Video Services Forum TR-06-1, resend a missing packet when the receiver reports it. For this to work, the configured latency has to be several times the round-trip time, so that a packet can be requested more than once.
Separating broadcast traffic
Put broadcast equipment in its own VLAN and, at larger venues, in its own routing domain (VRF). There are two reasons for this. One is operational: traffic from the guest Wi-Fi, the ticketing system or an office PC cannot compete with the broadcast streams in the same Layer 2 broadcast domain, and multicast flooding does not spread beyond the broadcast segment. The other is security: nobody on the audience network should be able to reach an encoder, a camera's web interface or a mixing console.
What the broadcast segment's rule set should cover
- Outbound from the broadcast segment, allow only the streaming platform's ingest addresses and the required time server, on the configured protocols.
- Keep device management interfaces on a separate management network that only the technicians can reach.
- Disable unused switch ports in public areas, or put them in a VLAN that forwards no traffic. At an event, the audience can also get to the network jacks on the walls.
- Change the devices' default passwords and turn off services you do not need.
The stream key, which the encoder uses to log in to the streaming platform, needs particular care. Anyone who knows it can broadcast anything on the channel in the organizer's name. It must not end up in a shared document or on a note someone could photograph, and it is worth changing after the event.
The article on segmenting the internal network covers the general aspects of segmentation.
Test broadcast and backup path
Faults in a broadcast network rarely show up during setup. A missing querier comes to light after a few minutes, and a line that is too narrow only under real load, when every service at the venue is sending traffic at once. Run the test with the full signal path and the real settings, ideally at the event location and at the same time of day.
- Have every encoder send at the bit rate used at the event for at least an hour. A shorter speed test will not show the peaks and the dropouts.
- Watch the outgoing signal on the streaming platform, on an external device that does not depend on the venue network. That way you see what the viewer sees.
- During the test, watch the switches' interface counters: dropped packets, CRC errors and QoS queue saturation.
- Test a failure: pull the cable of the primary uplink and check whether the broadcast fails over to the backup, how long it takes, and whether anyone has to step in manually.
- Write down the results: which settings worked, what differed from the plan, and who does what if the same thing happens live.
How to build a backup
For an internet uplink, the backup is a second line from a different provider, ideally over a different physical route. Many streaming platforms also accept a backup ingest: a second encoder, or the same encoder over another line, sends the signal in parallel, and the platform switches over if the primary drops out. The mobile network looks like a convenient backup, but at a sold-out event the audience's phones load the same cells, so it can slow down exactly when you need it.
Inside the venue, on professional production networks, SMPTE ST 2022-7 provides the redundancy: the source sends the same stream over two independent networks, and the receiver merges the two packet by packet. If a packet is lost on one path, the receiver fills it in from the other, with no switchover time. This does require two genuinely separate switch chains, powered from two separate circuits. An uninterruptible power supply for the broadcast switches, the encoders and the uplink router is just as basic a requirement. Without it, switching off one stage circuit takes the whole broadcast down.
The article on network monitoring covers which data shows, during a broadcast, where the traffic is getting stuck.
Related service: Broadcast and event networks →