jumbo frames

Thus spake Tony Hain:

Hence my original question. Packets across the GE will be 1500
unless you are packing them.

Or unless you actually have >1500 MTU to the hosts, which is quite
possible. A traffic study from MCI's backbone (obviously years ago)
showed nearly 40% (byte-wise) of their traffic was in packets >1500
bytes. With the death of FDDI, this has probably come down, but
GE-attached servers in colos should push it back up.

Assuming you are just passing the packets as received from the
aggregation switch, this would only happen if your router hardware
was better at managing jumbo buffer allocations than 1500B ones.
Clearly it will waste large chunks of memory, so do you have
measurements to show the actual performance increase?

Routers usually have separate buffer pools for common packet sizes (or
use buffer vectors), so the MTU of the interface does not noticeably
affect memory usage.

Router performance is, however, directly related to packet size, since
forwarding overhead is per-packet and not per-byte. It is much easier
to fill big pipes with 9000 byte packets than 1500 byte packets.



Thanks Stephen,

I understand all those points; I was reacting to the comment that the only
jumbo frames were between the routers, so the only ones coming in would be
1500. Look at John's notes from 4/25 8:10pm & 4/26 9:40am. There is no way I
know of for a larger frame by itself to increase the speed of a packet. What
was missing was the subsequent comment about tunneling those packets (thus
increasing the size beyond 1500) across the GE links to the distribution
router. Avoiding fragmentation clearly has performance gains, but
fragmentation is unnecessary if the packet size remains the same from end to
end. The original scenario was simply standard-size feeders into a router,
with jumbo frame GE between the core routers, then standard-size
distribution on the other end. The claim was that jumbo frames in the middle
made it go faster. If this is true I want to know how.