[address-policy-wg] Just say *NO* to PI space -- or how to make it lessdestructive

On Wed, 10 May 2006 17:53:36 +0900, Masataka Ohta wrote:
>Sometimes, pipelining can improve performance by a small constant
>factor with a small amount of additional hardware.

Sorry: *LOL*

With this statement ("small constant factor") you have just
proven that you are far away with your routing table lookup
theory from any real design.

With today's CPU's we are talking of about 20 pipeline
stages ...

The advances in processing speed in the last ten years
are - to a significant degree - based on advanced in multi-
level and parallel pipelined CPU architectures (called
superscalar processors).

>Pipelining requires pipeline resigters.
>
>If you have too much stages of pipelining, additional amount of
>pipeline registers kills the benefit.

Sorry, but it seems to me that you are talking about
developments you have never been involved in.

Pipeline registers are D-Flipflops, build as Master Slave
static or e.g. using transfer gates, they are the _basic_
elements of any real IC design. These elements
determinate - together with the conditional logic between
them - the maximum permitted clock frequency.
RAM is never ever faster than a single D-Flipflop
using the same technology.

Pipeline _stalls_ are a problem with CPU's, if we talk about
an conditional instruction and if the branch prediction fails.

With IP routing with multistage lookup there simply aren't
such conditional instructions.

IP routing fits ideal with pipelining, as there is no dependence
between different IP packets, the first section of the
pipeline decodes the first portion of the IP address, then
passes the packet header to the next section responsible
for the next portion of the address, at this time the first
section starts decoding the next packet etc.

If a delay is required, e.g. opening a new RAM page based
on the first decode, then the packet header is put in the loop
for one cycle.

This is the difference between _throughput_ and _latency_
required by a forwarding engine.

Thus if constructed the right way, handling reasonably
large tables should not be a problem at all.

Best Regards
Oliver Bartels



Oliver Bartels F+E + Bartels System GmbH + 85435 Erding, Germany
oliver at bartels.de + http://www.bartels.de + Tel. +49-8122-9729-0