SMPTE 2014: Uncompressed Video Over COTS Ethernet Switches

HOLLYWOOD—How many packets are dropped by Ethernet switches moving video? Thomas Edwards of Fox set about with Aperi Corp. to find out.

The short answer: “No packets were dropped. When the switch is properly used,” he said. “When you oversubscribe the Ethernet switches, you lose packets.”

Edwards proceeded to describe how the tests were conducted. IETF RFC 2544, the benchmark methodology for testing Ethernet networks, was employed, as well as RFC 3918, the methodology for IP multicast benchmarking. Snake tests were done for latency and packet-delay variation, or PDV.

A unicast universal protocol signal was transmitted between sending and receiving pairs. Throughput was equal to the number of test frames sent to it by the test equipment. Latency, frame loss, and longest interval back to black bursts was measured. Fully and partially meshed networks were tested.

RFC 3918 “had some limitations that bugged me,” Edwards said. “Trial lengths were not long… about 120 seconds. We like to see our television programs last more than two minutes.”

A stream analyzer was used to look for any dropped or re-ordered packets. Latency information based on timestamps was recorded. Measurements were based on a 6.4 nanosecond clock. Transmission/receive uncertainty was figured at plus or minus 13 nanoseconds to account for buffering, serializing and deserializing.

All switches tested were 10 GbEtop-of-rack, enterprise quality, and non-blocked at wire speed as advertised, Edwards said. All were 1 RU, with 48 ports, small form-factor pluggable plus, or SFP+, with the capacity to carry six HD 2022-6 10 GbE signals. Switches from six vendors were tested. None were named.

A “snake wiring configuration” was used for measuring packet delay variation. Every pair of ports was turned into a VLAN, each attached by copper cable, so that one flooded into the next. Every port then had the maximum amount of bandwidth going through it.

Edwards’ team did six streams per burst, around 809,000 packets per second for an hour. They looked at inter-packet time variance—the difference between expected arrival time in a stream and the time a packet actually arrives. By comparison, he said, latency is a quality of an individual packet—how long it takes for a packet to make it through a switch, based on timestamps in packets.

“Within about 1 microsecond of time, those packets always arrived,” he said.

In terms of latency, three of the switches tested out at less than two microseconds; the one with the most latency tested at around 3.5.

For fan-out testing, an additional switch was used for traffic synching. It was set up as one single VLAN. Traffic coming into port No. 1 went out via all ports. A physical cable unplugging/plugging test was done.

“There were no receive packet errors, drops or resequencing on monitored ports,” he said. “Same with the dynamic VLAN membership test… except for one switch controlled through an OpenFlow API.”

“Changing or deleting a rule that is currently being matched can result in a brief outage,” he said. “Up to 30 microseconds has been observed. You can add a higher priority rule, which seamlessly picks up packets and starts moving them. In OpenFlow, however, there are only 65,000 priority levels,” which at some point will be topped out.

Then the group asked, “what if you send actual 2022-6 streams?”

One switch was tested with a 48-port snake configuration and 22 hours of six, SMPTE 2022-6 720p59.94 streams. Another test introduced 30 hops through all six switches.

“There was no packet loss,” Edwards said. “Packet delay variation and latency.. there was nothing unusual.”

Packet delay variation was less than a 1.6 microsecond spread. Latency was less than 4 microseconds per switch hop. Outputs appeared to be immune to physical insertion and removal of SFPs and VLAN membership changes, except when using OpenFlow for SDN.

“Enterprise-quality commercial off-the-shelf 10 GbE top-of-rack switches can carry packetized video at 9.36 Gbps for long periods of time with zero packet loss,” Edwards concluded.

More testing will be done, he said.