Craig Partridge wrote:
the notion MPLS is faster to switch than IP reflects a poor knowledge of router innards, or a poor router design.
David Charlap then wrote:
...
Over time, however, this advantage disappeared. IP best-match lookup chips were developed that could do a proper IP lookup at full line rate for OC-48 and even OC-192. With line-rate IP lookup, it's no longer possible for something else to be faster.
An aspect of MPLS routing that this may be overlooking is the fact that MPLS tunnels can be designed with certain attributes which implement a form of policy routing that normal least cost routing does not implement. That is, blue packets can route over this link but not red packets unless there is an outage of the S.F to N.Y. Link. Another policy might say that traffic from this customer may exit the backbone at only a few points. These engineered routing decisions are configured into an MPLS network so these routing decisions are not made on a packet by packet basis. Once data is stuffed into an MPLS tunnel that was set up with certain policies in mind you know that the data comes out at the other end. Intermediate nodes don't need to be convinced to "non-optimally" route this data but only this data. Walt
In message <200108072211.WAA10474@i-14.isi.edu>, Walter Prue writes:
An aspect of MPLS routing that this may be overlooking is the fact that MPLS tunnels can be designed with certain attributes which implement a form of policy routing that normal least cost routing does not implement. That is, blue packets can route over this link but not red packets unless there is an outage of the S.F to N.Y. Link. Another policy might say that traffic from this customer may exit the backbone at only a few points. These engineered routing decisions are configured into an MPLS network so these routing decisions are not made on a packet by packet basis. Once data is stuffed into an MPLS tunnel that was set up with certain policies in mind you know that the data comes out at the other end. Intermediate nodes don't need to be convinced to "non-optimally" route this data but only this data.
Hi Walt: There are three separate issues (at least) here, so let's tease them out: * Current routing protocols don't do policy. Very right and a known defect in IP routing (though in part, they don't do it because in the general case, policy is hard) * Per hop policy decisions can be made more effectively in MPLS than in IP. Not true in theory unless you want to look very deep in the packet to identify the policy association, though it may be true in practice on certain current systems. * Instantiation of per-hop policy information via MPLS is more scalable than it would be in IP (not quite said above but an implied issue). Almost certainly not true (see above about general policy being hard being why IP doesn't do it). Craig
On Tue, Aug 07, 2001 at 06:55:09PM -0400, Craig Partridge apparently wrote:
There are three separate issues (at least) here, so let's tease them out:
* Current routing protocols don't do policy. Very right and a known defect in IP routing (though in part, they don't do it because in the general case, policy is hard)
And policy-based routing everywhere is not scalable. OK, we could argue about the future, but I suspect that no matter how much power we give router owners, they'll come up with policies that use it all.
* Per hop policy decisions can be made more effectively in MPLS than in IP. Not true in theory unless you want to look very deep in the packet to identify the policy association, though it may be true in practice on certain current systems.
MPLS doesn't require per-hop policy decisions. Policy decisions only need to be made at the edge, re FEC inclusion. Intelligence at the edge etc. Parallels with the diffserv model of classifying & marking packets at the edge so you only need to look at PHBs in the middle.
* Instantiation of per-hop policy information via MPLS is more scalable than it would be in IP (not quite said above but an implied issue). Almost certainly not true (see above about general policy being hard being why IP doesn't do it).
Instantiation of per-hop policy in MPLS consists of forwarding by LSP, except at the edge router. ..Scott (at the IETF)
In message <20010808103821.H1592@SBRIM-W2K>, Scott Brim writes:
* Per hop policy decisions can be made more effectively in MPLS than in IP. Not true in theory unless you want to look very deep in the packet to identify the policy association, though it may be true in practice on certain current systems.
MPLS doesn't require per-hop policy decisions. Policy decisions only need to be made at the edge, re FEC inclusion. Intelligence at the edge etc. Parallels with the diffserv model of classifying & marking packets at the edge so you only need to look at PHBs in the middle.
Hi Scott: Sorry I was too cryptic here -- sure MPLS makes a policy decision -- it decides how to forwarding based on the tag (e.g. the policy is embedded in the tag). My point is that you could just as easily associate the forwarding rule with a key, made up, say from source and destination address (which in some route lookup schemes requires only one more memory access than looking up purely on destination).
* Instantiation of per-hop policy information via MPLS is more scalable than it would be in IP (not quite said above but an implied issue). Almost certainly not true (see above about general policy being hard being why IP doesn't do it).
Instantiation of per-hop policy in MPLS consists of forwarding by LSP, except at the edge router.
Except that something has to decide where the the path goes (and thus, has to execute the policy at something close to a network wide level in terms of analyzing the network and instantiating the path). If you're suggesting we can do policy purely at the edges, then presumably a routing protocol could equally well force its policy information to only be computed at the edges. Yes? Or am I missing something? Craig
participants (3)
-
Craig Partridge -
Scott Brim -
Walter Prue