On Aug 9, 2006, at 4:12 PM, Erblichs wrote:
Roch Guerin,
Let me try to be abit more succinct. which restates
the below..
If one major component of cost is the time to
complete a task.
Then the cost or time is significantly greater
for the routing table than the forwarding table.
The biggest reason for this is because the forwarding
tables need faster memory. If we assume that forwarding
table memory is 4x faster than routing table memory,
then if we complete a forwarding table memory task 2x as
fast and keep the other constant, we only have about a
10% speeedup,
' however, if we double the speed of a routing table task,
than we can improve by about 40%.
This is all *extremely* dependent on hardware architecture, and
further makes the assumption that the largest component of time is
the memory access speed, which is very unlikely. For the
architectures I'm familiar with, updating the forwarding table is
*much* more expensive than tinkering with internal OSPF stuff, and
that has *nothing* to do with memory speed.
Thus, improving the speed of the equiv of SPF processing
gives us a decent improvement, which leads to my disagreement.
Yes, this is Amdahls Law..
However, if we measure all of the tasks that are needed
for a rec'ved LSA over time, most are LSA age updates, which
removes the need for SPF calcs, and thus increases the
chance that a improvent elsewhere would be more beneficial.
You're further making an assumption that reducing the overall
instruction count over time is (a) beneficial, and (b) worth adding
system complexity for. Both are debatable and lie at the heart of
system design.
You seem to be making a common mistake, which is looking at a small
piece of the puzzle and trying to optimize that, while ignoring the
cost of doing so in reliability and stability. Taking your earlier
example, where you wish to switch the forwarding table as fast as
possible in the face of a failure, you have to take a look at the
entire time budget between the event causing the problem (the backhoe
cuts the fiber) and the restoral time. The cost of doing a full SPF
is likely to be a small fraction of this process, and the gains in
optimizing the process will be a small fraction of that small
fraction. If you can get the time budget for all the rest of this so
that it doesn't overwhelm the cost of a full SPF versus some kind of
shortcut, then it's time to start talking about gains in route
calcuation.
--Dave
_______________________________________________
OSPF mailing list
[email protected]
https://www1.ietf.org/mailman/listinfo/ospf
