Imagine, if you dare, if Ethernet’s maximum frame size was merely 15 bytes. That’s 120 bits out of the 10 million bits per second the medium can transport. What a nightmare that would be.
It’s too bad the nightmare is a reality. Relative to the line rate, that is what Ethernet’s 1,500-byte maximum frame size looks like in a gigabit Ethernet network. Vendors say the small frame size is not a problem, but it is — a big one.
Gigabit Ethernet vendors admit that today this puny maximum frame size presents a problem. You see, even in a ridiculously optimistic scenario, where a session streams nothing but maximum-size frames, it takes 80,000 such frames to fill the pipe. Vendors admit that even a multiprocessor server trying to field this many interrupts per second would be brought to its knees.
So far, everyone agrees. It is the discussion of the solution to this problem that engenders discord.
A majority of Ethernet switch and network interface card (NIC) vendors suggest the real problem is the frame processing burden put on the server CPU. They claim the answer to the problem will be soon-to-be-available NICs with onboard CPUs to which today’s server-based network functions can be offloaded. The vendors confidently maintain that as soon as the new NICs are rolled out the problem will be solved.
The problem won’t be solved; it will simply be relocated. Now instead of the server having to waste processing power handling ridiculously small frames, the CPU on the NIC gets to waste power doing the same job.
CPUs and buffer memory cost money — wherever they reside. Why vendors maintain that moving the waste to a different location is a good fix is beyond me. Furthermore, these server-offload NICs won’t come cheap.
Why won’t vendors embrace large frames? There is enormous resistance to changing the Ethernet standard to accommodate a larger frame size. The vendors claim it would cause enormous backward-compatibility problems and make every piece of Ethernet gear already deployed obsolete.
Excuse me? Couldn’t this same argument be used against fast Ethernet? Aren’t standard 10Mbps adapters incompatible with fast Ethernet? Of course they are — the adapters don’t understand what fast Ethernet is.
Fast and gigabit Ethernet are simply supersets of the original Ethernet standard. Why can’t Jumbo Frames work the same way?
Network connections of 10/100/1,000Mbps are either configured manually or, more frequently, autodetected. The ports determine the level of support and communicate at the most sophisticated level supported by both. Can’t frames work the same way?
Token-ring uses this approach quite effectively. At session initiation, each hop marks the handshake frame with the maximum frame size it can support, and the endstations autoconfigure their session parameters not to exceed this value.
Ethernet vendors need to face this problem, stop being so rigid and let the Ethernet standard enter the gigabit era.