The MPLS WG Archive

[Curtis Villamizar <curtis@fictitious.org>]

In message <3DF6AF96.FD3EECB1@cad.zju.edu.cn>, Jing Shen writes:
> 
> 
> What I care about is how may those changes affect the performance. From
> your
> word I come to see that our experimental result may only stick to our
> implementation
> and our experimental policy, but I'm still not clear that how could a
> ASIC implemention
> of MPLS engine compares to a ASIC implementation IP engine? Is there any
> measurement data or paper on this?


Probably some work has been done but it doesn't make a whole lot of
difference.  The best modern ASICs forward at full line rate for
either MPLS or IP at OC192c rates of roughly 10Gb/s.  So figure a few
10s of nsec per table or route lookup.  This doesn't amount to a whole
lot of time compared to any normal speed of light delay for a real
network hop and even exceeds the packet serialization time.

For UDP or TCP, the difference is 4 bytes added to the packet for
MPLS.  For 1500 byte TCP MSS that's under 0.3%.  For 552 its about
0.7%.  For a 40 byte packet its 10%.  If you want traffic engineering
it certainly doen't have the overhead of ATM (9-62% overhead).

The network is usually not full so the <1% has negligible effect.  TCP
or UDP performance is typically dominated by window size and only
mildly impacted by transient queueing delays until the network fills,
then determined by drop characteristics after the network fills.

That's why this issue hasn't been a hot topic.

Curtis