Limiting NIC speed leads to better video quality

Question

I have two workstations A and B connected to the same L3 switch at 1Gbps. Workstation A generates an MPEG2-TS UDP video stream at 10Mbps data rate. Workstation B receives this stream correctly. The switch is connected to an edge router via 100Mbps port.

Workstation C, which is on the other end of a private WAN, receives poor quality video stream: the first N lines of the image are good quality (N randomly changes) and the other lines suffers severe pixel blocks. WAN link capacity is 20Mbps.

The fact the first N lines are always good quality made me think it is not random packet loss. However I didn't understand at all why a 10Mbps stream could saturate a 20Mbps link.

To my surprise, after limiting NIC on workstation A to 100Mbps, workstation C could correctly receive and display the video stream. Why?

Answer 1

The likely cause of this is small buffers in the switch combined with poor design of the video streaming application.

Most likely the video streaming application doesn't send a paced stream of packets but instead passes a whole frame worth of packets to the OS at once. The buffers in the client OS and the edge router are big enough to handle this but the buffers in the switch are not.

Workstation A sends that burst of packets to the switch at 1Gbps but the switch can only send it to the edge router at 100MBps. The buffer in the switch is too small to handle the burst and so packets get dropped.

Dropping the link to 100Mbps on workstation A means that packets no longer pile up at the switch because the input and output ports are running at the same speed.

You should be able to confirm this by looking at frame loss counters on the switch.

Answer 2

There are several possibilities:

1. Workstation C's cable causes errors = frame drops at gigabit speed or with all twisted pairs in use. Check the workstation's error counter at full speed under load. You can check either side for 1000BASE-T as pairs are used bidirectionally.

2. There is a congestion or buffering problem on the way to workstation C. Check routers and switches in between for errors/congestion.

3. There is a setup problem (jumbo frames, MTU mismatch or misdetection).