[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
AW: [PWE3] PWE3 question
Title: RE: [PWE3] PWE3 question
Let me
add some precise numbers computed based on an example network with 300 nodes and
550 links. n=100 nodes are PEs, 200 nodes are Ps.
The
PEs are distributed randomly but fair. And so is the topology as
well.
Case
1: Have n times (n-1) p2p LSPs, i.e. 9900 p2p LSPs, each routed according to the
"shortest path" principle.
Computation result: 99874 hops(=links) are needed.
Case
2: Have n mp2p LSPs, i.e. 100 mp2p LSPs, each routed according to the "shortest
path to the egress" principle.Alltogether they take 19886
hops(=links).
Case
3: Have n mp2p LSPs, i.e. 100 mp2p LSPs, of identical smallest-size tree
structure (each time another PE is the root of this tree).
100 times 160 hops, i.e. 16000 hops are needed.
Case
4: Have n-1 pairs of mutual inverse LSPs which form a bi-directional tree of
(elementary) tunnels between the n PEs. They are selected such, that the number
of
hops is minimal. 2 times 169 hops = 338 hops are needed. Additional
Hierarchical LSPs, i.e. concatenations of the elementary
LSPs, shall take care
that there is a tunnel sequence
from any ingress PE to any egress PE.
a) [n * (n-1) ] - 2*(n-1) non-merging
LSP-sequences (=p2p H-LSPs) are needed
or b) n merging (mp2p) H-LSPs are needed
Allthough b) is more conservative on resource consumption, you may consider the
cost for the H-LSPs according to a) as well according to b to be "ZERO
for the network" (see may
draft-hummel-mpls-hierarchical-lsp-00.txt)
Case 1
is indeed the n-suqare problem (99873 hops). Merging LSPs (see case 2
and case 3) reduce the number of hops considerable, let me say by 80
%.
Case 4
however is the way to overcome the n-square problem (338 hops instead of 99874
hops, i.e. a reduction by 99.66 %).
Will
say: Merging LSPs do no overcome the n-square problem. I am sure, If I extended
the network to let's say 1000 PEs and have similar meshes ( 1 to
5
neighbor nodes per node), the reduction due to merging
LSPs stays by 80 % whereas the H-LSPs would yield 99.9966 %.
As you
can see, the saving due to H-LSPs are immense.Therefore, why shouldn't we
concatenate elementary LSPs, i.e. build H-LSPs?!
It is
the same logic why we build routers instead of having a full mesh of physical
links in the internet!
There
are more arguments:
-
Neither the n*(n-1) p2p LSPs nor the n mp2p LSPs can support Multicast at
all (i.e.enable that some packet doesn't have to be sent multiple times over the
same physical interface).H-LSPs however will reduce that this
will happen at least considerably. I really wonder why it is acceptable that a
P-router may deal with
VPN-multicast, and why multicast blind traffic is
acceptable (towards PEs which have no receivers for that Multicast application
instance).
- One
can hardly affort x times n*(n-1) p2p LSPs or x times n mp2p LSPs, as to support
alternate routig (traffic balancing, QoS-/Policy-sensitive
routing).
But one can easily affort x times the number of H-LSPs. The costs for the
network: x times ZERO = ZERO. Indeed, these costs are ZERO, when
the
messages for establishing the H-LSPs are tunneled thru these elementary LSPs,
i.e. are completely invisible to the core network.
Regards,
Heinrich Hummel
Siemens AG
Eric:
This one has been nagging me for a while:
<snip>
> Case 1: Martini signaling, mp2p LSPs. The amount of
> resources needed in each
> P router is O(n).
An mp2p LSP mesh can be said to have O(n) egress points, but
it has O(n**2) ingress points. So if I compare the actual fabric
consumed between a full mp2p and a p2p mesh I would get O(n**2) * (average
network diameter) LSHs for p2p, and some fraction of this for mp2p (again a
function of diameter in hops reflecting on how fast the ingresses merged on
the egress). mp2p requiring uni-directionality then doubles the fraction as
far as actual administrative/signalling load is concerned (each direction
requires a separate set of transactions to establish).
So one can claim some incremental improvement in scalability
via the use of mp2p but it is not this exponential thing that everyone waves
about......in other words single level MPLS is a small fixed amount of time
apart from ATM on Moores law curve (stacking being a different
story).
cheers
Dave