Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: Jumbo Frames in 10GbE?

> >Indeed, there have been some pleas for "no superpackets, please!" on this
> >reflector.
> Yes.  The technical rationales supporting these pleas would be interesting.

We go through this exercise every three years, give or take. Long packets are
efficient at the endpoints, where burstiness is good. Less overhead per unit
of transfer. Burstiness is bad in the cloud, specially for packets that have
to sit in queues behind these long packets. Burstiness is sometimes painful for
managing buffering resources (whenever memory has to allocated proportionally
to packets, like the XOFF RTT input buffer for 802.3x).

Note that the processing burden for devices in the cloud is determined by 
the short packet case (64 bytes), and to the best of my knowledge jumbo 
frames do not provide any relief there.

> At 10 gigabit, what existing base is there to protect?  One would have
> thought there was none.

I believe that one of the powerful features behind the 100x speedup of
products based on the 802.3/802.1 duo (in less than 10 years) was the
interoperability based on link speed independence. That interoperability
with 802.3 devices is there and should be protected.

Furthermore, the adoption of higher speeds has traditionally started
from traffic aggregation, and the WAN appeal of 10Gig makes this more
applicable than ever. Aggregating slower sources into 10Gig, while using
MTUs larger than 802.3, means that the sources were not 802.3 to start

Ariel Hendel
Sun Microsystems