What are limiting factors of Ethernet payload size regarding the future of internet traffic?

Question

The ethernet payload is 1500 bytes. Devices have to deal with a huge number of frames, where many are fragments of one and the same message. More than a thousand just for an image file for example.

Data transfer could rise more and more in the future, with millions or much more fragments for a message during its life-time on its way from A to B.

So, I wonder if extending the frame size is an option, with the idea of getting the same done as current ethernet but with less frames and less fragmentation overhead.

So far, I only thought of these potential problems:

• the CRC may fail often when the frame is too big
• big frames could be less flexible regarding flow control

Answer 1

We've had jumbo frames for decades.

One reason they're not in common use on the Internet is IPv4 (and later IPv6) overlooked any possibility of having underlying layer-2 networks with connected nodes having different MTUs (or MRUs.)

ARP and ICMP do not have the ability to allow nodes to negotiate a mutually-agreeable MTU/MRU.

So for example, Internet Exchange Points (IXPs) have dozens or hundreds of connected members, and they all generally agree on 1500 bytes MTU because any larger size would have to be agreed on by everyone on that IXP, otherwise, the result would be black-holing behavior; some parts of the Internet couldn't reach others -- but only for the particular packets exceeding the MTU/MRU of one member in a connection/session. Very problematic!

The only major IXPs which have a broad base of jumbo frame support are Netnod and SIX. Even among these, there's not wide agreement over what the larger MTU should be -- some Netnod exchanges offer 4470 (SONET/SDH-related size) and others 9000. That's disagreement just within one operating organization that runs IXPs!

The benefits of jumbo frames over the Internet are fairly small and the complexity of effectively utilizing them remains non-trivial.

Answer 2

You’re missing the biggest problem: billions of dollars of equipment would need to be replaced, and essentially all at once. Who would pay that for just a modest increase in efficiency?

In your previous example, you'd save the 14 bytes per Ethernet frame, or less than 1% of the total transmission size. Sounds like a hard sell to me.

Answer 3

In addition to Ron's answer:

So far, I only thought of these potential problems:

• the CRC may fail often when the frame is too big
• big frames could be less flexible regarding flow control

Absolutely. In addition, larger frames can cause more latency/jitter in spite of QoS - larger frames means less granular traffic.

Since frames are handled by hardware, processing fewer or more isn't really a big difference. The improvement in effiency isn't a big thing either:

• TCP over IPv4 over Ethernet with MTU 1500: 1460 bytes payload with 1538 bytes on the wire = 94.93% efficiency
• same with MTU 9000 jumbos: 8960 bytes payload with 9038 bytes on the wire = 99.14% efficiency

Compare the cost of replacing all the hardware with the effiency gain of 4% and you've got your answer. You can simply increase the available bandwidth tenfold for less cost.