IPv6 day and tunnels
mohta at necom830.hpcl.titech.ac.jp
Tue Jun 12 15:33:57 CDT 2012
Templin, Fred L wrote:
>> Have you learned enough about Moore's law that, at 10Gbps
>> era, 72us of delay is often significant?
> I frankly haven't thought about it any further.
That's your problem.
> You say
> 1280+ belongs in ITU, and I say 1280- belongs in ATM.
As I already said, 9KB is fine for me.
Small cell size (32~48B, not 1280-) of ATM is derived from
slow (64Kbps voice) speed and short (0.1s) delay requirement
with fair queuing and has no relevance to today's network.
> Larger packets means fewer interrupts and fewer packets
> in flight, which is good Moore's law or no.
That is a basic misunderstanding of those who thought
jumbograms were good.
They (or you) thought supercomputers are vector computers,
very slow to react against interrupts, and have no IO
processors to take care of packet handling.
The reality with Moore's law, however, is that NIC cards
can take care of even TCP, which makes jumbograms totally
Moreover, the huge number of scalar processors in modern
supercomputers means communication granularity is (depending
on computation algorithm) often tiny, which means networks in
supercomputers must be able to handle small packets efficiently.
Larger packets means, in addition to longer HOL
blocking, more delay to pack more data in the packets,
even though processors often want to receive data with
Thus, as with other features of IPv6, jumbograms are no useful
More information about the NANOG