Hi Peter,
Please see inline:
*From:*Peter Psenak <[email protected]>
*Sent:* Thursday, March 21, 2024 5:39 PM
*To:* Dongjie (Jimmy) <[email protected]>; Les
Ginsberg (ginsberg) <[email protected]>; [email protected];
[email protected]
*Subject:* Re: [Lsr] Comments on draft-zhu-lsr-isis-sr-vtn-flexalgo-07
Hi Jie,
On 21/03/2024 02:34, Dongjie (Jimmy) wrote:
Hi Les,
Thanks for providing your opinion with an example.
In your example, the default IGP metric is used, which is normally
calculated based on bandwidth. While Flex-Algo can support metric
types such as TE metric and delay.
When Flex-Algo is used as the control plane of NRP, usually the
metric types other than IGP metric would be used. We could add
some notes about the selection of metric type to this document. In
most cases per Flex-Algo metric type would not be needed.
Your proposal of making each member link an L3 link is an
alternative solution, while that would bring back the problems we
discussed during the L2 bundle standardization, and can impact the
network stability and scalability.
Your second proposal (controller based path computation and
construction) works for scenarios where strict TE-path SID-list is
used to steer traffic into specific bundle member links on each
hop, while traffic with Flex-Algo prefix SIDs will be mixed up and
ECMP among the member links of the L3 interface.
So we do see there is a gap in using Flex-Algo to support NRP, and
would like to hear feedbacks from the WG on possible solutions
(including this one).
there is no gap in Flex-algo. Flex-algo is a routing concept and as
such only works on L3 constructs. That will not change.
*[Jie] Fully agree Flex-Algo is a routing concept and works on L3
control plane, while it shows a gap in how to map Flex-Algos to
different subset of resources for network slicing. Currently traffic
of different Flex-Algos would share the same set of resources on the
L3 outgoing interfaces. *
The problem is that you are trying to mix the routing (flex-algo) with
the PHB/QOS. These are two different things.
You can achieve PHP/QOS by marking the packet and give it a necessary
treatment you need at each hop, e.g. reserve certain bandwidth to it,
or even reserve a L2 bundle for it if that's what you want.
*[Jie] QoS PHB is per-hop behavior, which cannot provide end-to-end
resource guarantee at NRP/Slice granularity. Consider the difference
between DiffServ and IntServ. And within each NRP, QoS is still needed. *
*Reserving an L2 bundle member link for NRP is the approach proposed
in this document. *
Alternatively, you can classify the traffic at each hop using other
mechanisms, but it becomes slow and problematic. **
*[Jie] Agree that per-hop traffic classification has many problems. *
What you propose is to overwrite the routing decision and instead of
using the L3 outgoing interface computed based on L3 information, you
install the specific L2 bundle member out of such L3 interface in
forwatding. It works, because by using the L2 member of the L3
interface the traffic is forwarded to the same next-hop as has been
calculated by the L3 routing. Nobody can stop you doing that locally
if you wish doing so. But there is absolutely nothing you need from
the IETF to do this. There is no need to advertise anything to do what
you describe, as this is all a local behavior of the node. There is no
need to add a new E-bit, and there is not even a need to advertise
affinities for the L2 bundle members.
*[Jie] The distributed path computation is still based on the L3
links/interfaces, the change is in the forwarding entry installation.
Thanks for confirming it works. *
*The advertisement of the L2bundle information is for the controller
or ingress nodes to perform path computation based on NRP-specific
constraints and can use Algo-specific SIDs together with bundle member
Adj-SIDs in building the SID list, this aligns with the usage of L2
bundle and extends its applicability to Flex-Algo-specific SIDs. *
*The E bit is to indicate the L2 bundle is working in the exclusive
mode (rather than load balancing), which means the Flex-Algo SIDs can
be used to steer traffic to the corresponding member links. *
**
*Best regards,*
*Jie*
**
I see no need for this draft.
thanks,
Peter
Best regards,
Jie
*From:*Les Ginsberg (ginsberg) <[email protected]>
<mailto:[email protected]>
*Sent:* Thursday, March 21, 2024 10:36 AM
*To:* Dongjie (Jimmy) <[email protected]>
<mailto:[email protected]>; [email protected];
[email protected]
*Subject:* RE: Comments on draft-zhu-lsr-isis-sr-vtn-flexalgo-07
Jie -
Thanx for the quick response and confirming that my understanding
of the intent of the draft is correct.
Making a routing decision when the full topology information is
not provided as input to the Decision Process leads to incorrect
or sub-optimal routing. Here is one simple example.
Consider the following simple topology (Layer 3 links):
B
/ \
A D
\ /
C
All layer 3 links participate in Flex Algo 128.
On both B and C, the Layer 3 link to D is an L2 bundle and the
total bandwidth of the bundle links are the same.
On link B-D, the L2 bundle member assigned to the NRP associated
with flex algo 128 has 100 Mb of bandwidth.
On link C-D, the L2 bundle member assigned to the NRP associated
with flex algo 128 has 1 GB of bandwidth.
The L3 SPF associated with algo 128 utilizes Layer 3 metric
advertisements. Based on that, traffic from A to D will be equally
balanced via B and C.
However, what you intend is that when algo 128 traffic is
forwarded by B it will utilize a 100 Mb link – whereas when algo
128 traffic is forwarded by C it will utilize a 1 Gb link.
Clearly the ECMP traffic flow which is the output of the L3 SPF is
sub-optimal.
How could this be fixed?
1)Do not use L2 bundles on B and C. Make each bundle member an L3
link and run IS-IS on the Layer 3 interfaces. In such a case
different L3 metrics can be advertised for each L3 link and Flex
Algo 128 can be associated only with the desired L3 link on C and D.
Standard flex-algo as defined in RFC 9350 works and requires no
modifications.
2)Do not use L3 routing/flex algo. Define some other mechanism to
mark packets in a way that the forwarding recognizes as mapping to
the appropriate L2 link.
The L2 bundle advertisements provided by IS-IS as per RFC 8668 can
be used by this (external to IS-IS) mechanism.
For example this mechanism could use the admin group advertised
for each L2Bundle member to determine the mapping between an NRP
and a link.
All of the functionality required is already defined in RFC 8668 –
the only thing you need to define is this new mechanism – which is
not part of IS-IS and therefore does not belong in an LSR draft.
NOTE: Please do not suggest that a different metric-type can be
used for each Flex-Algo. Such an approach does not scale as it
requires as many metric-types as Flex-Algos – which we do not have. 😊
What you MUST NOT do is use L3 routing to make a routing decision
for a topology which is not part of the input to the routing
decision process. But that is exactly what you are proposing in
this draft.
Hope this example is clear.
As regards the clarity of Section 4, that section simply says
(using the SR-MPLS text):
“A forwarding entry MUST be installed in the forwarding plane
using the MPLS label that corresponds to the Prefix-SID associated
with the Flex-algorithm corresponding to the NRP.”
But this entry must have next hops which include only the L2 links
associated with the NRP mapped to Flex-algo 128. How this is done
is not described – but as it requires using information advertised
in the L2 Bundle Member Descriptors this clearly cannot be done by
IS-IS w/o violating RFC 8668. IS-IS will simply attempt to install
a forwarding entry based on the L3 topology – which will indicate
to use the L3 link. How this forwarding entry is
discarded/overwritten is not specified. But, this is a problem
which should never need to be solved.
Les
> -----Original Message-----
> From: Dongjie (Jimmy) <[email protected]>
> Sent: Wednesday, March 20, 2024 4:30 PM
> To: Les Ginsberg (ginsberg) <[email protected]>; [email protected];
draft-zhu-lsr-
> [email protected]
> Subject: RE: Comments on draft-zhu-lsr-isis-sr-vtn-flexalgo-07
>
> Hi Les,
>
> Thanks for the review and comments.
>
> Please see some replies inline:
>
> > -----Original Message-----
> > From: Les Ginsberg (ginsberg) <[email protected]
<mailto:[email protected]>>
> > Sent: Thursday, March 21, 2024 7:32 AM
> > To: [email protected] <mailto:[email protected]>;
[email protected]
<mailto:[email protected]>
> > Subject: Comments on draft-zhu-lsr-isis-sr-vtn-flexalgo-07
> >
> > This draft discusses how to use flex-algo in support of Network Resource
> > Partitions (NRPs). In particular, it proposes to use a
combination of L3 links
> and
> > L2 Bundle member links as the topology associated with a given NRP. In
> those
> > cases where an L3 link is using an L2 bundle and individual bundle
members
> > are "assigned" to different NRPs, it then proposes to associate the
parent L3
> > link with multiple flex-algos. The intent seems to be to utilize the
L3 algo
> > specific SIDs to assign the traffic to subsets of the L2 Bundle
members.
>
> Your reading of the intent of this document is correct.
>
> With the proposed mechanism, traffic with Flex-Algo specific SIDs could be
> steered to different partitions of the L3 link resources.
>
> The only thing I'd like to mention is the L2 bundle members could be
virtual or
> physical, they are just used to represent different subsets of the link
resources.
>
>
> > This is extremely problematic.
> >
> > The output of the L3 algo-specific SPF will be to install nexthops
pointing to
> the
> > L3 interface for packets which arrive with the L3 algo specific SID.
But since
> the
> > intent is to only forward traffic for a given algo specific SID
via specific L2
> > Bundle members, the L3 forwarding entries will have to be
overwritten - in a
> > manner not specified by the draft.
>
> Section 4 of this document specifies the required forwarding plane
behavior
> and the forwarding entry installation.
>
>
> > The implementation complexities this introduces arise because the
proposed
> > solution attempts to use a Layer 3 technology (flex-algo) to
control the use
> of
> > L2 links. This should not be done.
>
> In the proposed mechanism, Flex-Algo is still used for constraint path
> computation, and only the L3 links attributes are used in the
computation. The
> L2 member links are only to partition the resources used by different
Flex-Algo
> traffic.
>
>
> > Indeed, even independent of flex-algo, trying to use a Layer 3
routing
> protocol
> > to control traffic flow on an L2 sub-topology is broken.
> > It means that the L2 bundles have been improperly defined for use by
the L3
> > routing protocol.
>
> There is no routing computation based on the "L2 sub-topology", as L2
bundle
> member links are not visible in the L3DB. All the Flex-Algo computation is
> based on the attributes of L3 links.
>
>
> > RFC 8668 defines the advertisements of L2 Bundle member link
attributes by
> > IS-IS. The introduction of RFC 8668 states:
> >
> > "...the new advertisements defined in this document are intended
to be
> > provided to external (to IS-IS) entities."
> >
> > This means these advertisements are not to be used by the routing
protocol
> > itself. The association of these advertisements with the Layer 3
SIDs defined
> by
> > Flex-Algo is a clear violation of the intended use as stated by
RFC 8668.
>
> As stated above, L2 bundle link attributes are not used in path
computation.
> The Flex-Algo specific SIDs still point to the L3 interface based on that
> computation. The only change is that a Flex-Algo SID can further
points to the
> resource of an L2 member link (consider it as a subset of the resource of
the L3
> link if that is easier to understand). So the L2 bundle information is
only used
> for associating different Flex-Algo SIDs with different subsets of
resources of a
> l3 link.
>
>
> > This draft should be abandoned.
> >
> > NOTE: None of the points above should be interpreted to mean that flex-
> algo
> > cannot be used in support of NRPs. (Whether that is a good idea
or not is
> out
> > of scope for this discussion).
>
> AFAIK people are talking about using Flex-Algo to support NRPs. This
> document provides a solution to meet their needs.
>
>
> > But the proper way to do that is when the NRP maps to an L3 topology.
Such
> a
> > usage is fully supported by RFC 9350 and there is no need to write an
> > additional document to define how this is to be done.
>
> In some cases it is possible to map different NRPs to non-overlapping L3
sub-
> topologies, while in many other cases the same L3 link needs to
participate in
> multiple NRPs, each of which is assigned with a subset of the link
resources.
> The latter case cannot be supported by RFC 9350, and it is the target of
this
> document.
>
>
> > In cases where an NRP maps to an L2 topology, some other mechanism
> needs
> > to be defined as to how forwarding entries for a given NRP are
determined
> and
> > installed. Such a mechanism would qualify as "external to IS-IS" and
> therefore
> > could make use of RFC 8668 advertisements.
>
> This document also provides descriptions about this. As I mentioned it
is after
> L3 computation, and makes use of the L2 bundle information.
>
>
> > But attempts to utilize the Layer 3 Flex-Algo technology to control
traffic flow
> > in an L2 topology are misguided and flawed.
>
> As long as Flex-Algo is used for L3 topology based computation, IMO it
still
> complies to RFC 9350.
>
> Best regards,
> Jie (on behalf of coauthors)
>
> >
> > Les
_______________________________________________
Lsr mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/lsr
