[ppml] 2005-1 and/or Multi6
Michael.Dillon at radianz.com
Michael.Dillon at radianz.com
Fri Apr 15 12:28:45 EDT 2005
> > At the end of the day, it is possible to optimize by
> > processing fewer bits
> This would have been true in the past, but not anymore. By the time the
> size of the routing table is so big that it requires optimization, there
> will be no processor that does not have a 64-bit core and processing
> 64-bit elements will be as fast as processing 32-bit elements.
This may be true in the domain of trendy laptop and
desktop computers, but I cannot believe that it will
be true overall. We still have millions of 8-bit CPUs
being built into devices every year and there is nothing
on the horizon that indicates manufacturers will stop
using them. Nowadays, these 8-bit CPUs are usually called
PICs and they do jobs like manage buttons, blinking LEDs,
motor control, etc. Next up is the FPGA, which is
programmable at the hardware level, i.e. you can tell
the device how to organize it's logic gates internally
to change from being an 8086 to an ARM 7 to a 68020 to
a special purpose neural network machine. In a world
where reconfigurable hardware elements become faster and
more powerful, the concept of "the CPU" begins to falter.
If there is a speed advantage to processing 32 bit elements
then there will be hardware available to do that job
for the next few centuries. Ditto for variable length
elements. People do not design and produce these things
to follow fashion trends; they design them to do real
work and meet real demands.
> Besides, do some basic math:
> 1 million IPv6 routes/paths (we're not there yet, are we?)
> Stored at 32 bits: 4 Megabytes
> Stored at 48 bits: 6 Megabytes
That is arithmetic not math. If you want to do some
math you would include the overhead of a trie,
the average number of interfaces on a router (after all,
routes point to interfaces) and the number of views
that the average core router would need to carry.
And a real analysis would also look at gate counts
for a hardware implementation with numbers for various
implementations of TCAM (Ternary Content Addressable
Memory) as well as other forms of hardware encoded
> (the other associated storage such as the AS-PATH would not change)
> Difference: 2 Megabytes. Give me a break, this is not worth any code
> complexity nor custom chip design even as of today.
Maybe you should have a discussion with somebody who
actually designs routers. Or just look inside a router
and see if you can identify every one of the chips as
a standard non-custom chip. FPGAs do not count as standard
> > If an organisation has an AS number then their
> > network is not a "site".
> Where does this come from?
That comes from way back at the beginning of the
thread and the policy proposal that started it all.
Should ARIN give an IPv6 PI allocation to any
organization with an AS number. I am saying, Yes ARIN
should do so and that PI allocation should be a /32
for reasons of prudence and conservation of future
More information about the ARIN-PPML