Re: [bess] VXLAN BGP EVPN Question

["Rabadan, Jorge (Nokia - US/Mountain View)" <[e-mail]>]

Hi Gyan,

Actually we started with the evpn dci draft in 2013 :-)

The way I see the unknown mac route it saves flooding if all the MACs in the POD/DC are known beforehand. The unknown unicast traffic can be aliased to the GWs. In case of failure in one of the GWs, the AD per-ES route for the I-ES will be withdrawn (mass withdraw for all EVIs) and the unknown traffic can be sent to the redundant GWs. So this failure won’t generate any extra flooding.

Thanks.
Jorge

From: Gyan Mishra <[e-mail]>
Date: Saturday, April 25, 2020 at 8:45 AM
To: "Lukas Krattiger (lkrattig)" <[e-mail]>, "[e-mail]" <[e-mail]>
Cc: BESS <[e-mail]>, Jeff Tantsura <[e-mail]>, "Rabadan, Jorge (Nokia - US/Mountain View)" <[e-mail]>
Subject: Re: [bess] VXLAN BGP EVPN Question


+ Ali

Lukas

I noticed that Ali was on the multi site draft which I which expired in 2017 around the same time the DCI overlay  draft was submitted.  I went through the logs but did not go through the mail archives to see what happen to multi site draft.  My guess is these were two competing drafts and multi site was geared solely to EVPN procedures for vxlan encapsulation and thus did not achieve WG adoption, where your DCI overlay draft accounts for every encapsulation type using EVPN procedures and is more comprehensive approach to DCI providing an improved solution to Multisite vxlan overlay stitching.

I like the re-origination of the VNI and RD idea using local context on the gateway as an additional control mechanism which prevents Type 2 mac-ip routes from being flooded between pods that should not without flood filters. With the multi site feature there are no control and all mobility routes are flooded unfortunately active or not.

With this draft is it possible to add a feature for conversation learning of only active flows when the type 1 BGP a-d is sent for initial BUM advertisement for arp or nd, there could be a snooping mechanism similar to IGMP snooping that discovers the active flow and thus creates the control plane level type 2 Mac-IP state followed by being flooded in data plane NVE tunnel overlay.  I think this concept could apply intra site fabric leaf to leaf but I think would be extremely beneficial for inter pod or inter site.

This could be separate feature or option to the selective advertisement.

So the selective advertisement works in conjunction with re-origination of RD and locally significant VNI.

So what I would envision with the conversation learning active flow detection feature you would use global VNI and now only the active type-2 Mac-IP routes would be propagated inter pod or site.

This feature would be a tremendous benefit to operators and help with mac scale.

In our Cisco multisite feature implementations we do use the recommended BUM traffic multi site feature specific suppression applied on the BGW.  So that definitely helps with the BUM suppression for sure.

In section 3.5.1 UMR - so the route type is like a default Mac route 0/48 with ESI set to DCI gateway I-ESI for all active multi homing, and so instead of flooding all mac’s and have to rely on mass mac withdrawals during a failure, now only the UMR is withdrawn.  Is that correct?

That’s a huge savings on resources.

Kind regards

Gyan

On Fri, Apr 24, 2020 at 3:25 PM Lukas Krattiger (lkrattig) <[e-mail]<mailto:[e-mail]>> wrote:
Thanks Jorge and Jeff for guiding all the way thru the features and functions we have around, in DCI-overlay and Multi-Site.

Gyan,

Specific to the VNI distribution, BUM handling and the re-origination in Multi-Site.
With re-origination, the RDs are changed on the GW node. With this in mind, the VNI could be Global or local significant. In the case of local significants, we can stitch VNIs together (ie (VNI1 - GW - VNI2 - GW - VNI3).
Further, MAC- or IP-VRFs that are not supposed to be extended to a remote Sites will not advertise any MAC or IP routes beyond the local GW. This way you will keep the control-plane clean and avoid unnecessary creation of flood lists. This is what we call selective advertisement, which is different than conversational learning. Conversational learning could be a complement to selective advertisement. The unknown MAC approach that Jorge mentioned is a different approach for similar optimizations.
In addition to ARP suppression, in the specific Cisco implementation of Multi-Site, we provide a BUM traffic policer to rate limit between Sites. This policer are located on the GW and acts in the egress direction.

So with the DCI EVPN VNI translation does that end up netting the desired effect control plane segregation from data plane and providing that reduced size Mac VRF showing only active interesting traffic type 2 Mac-IP routes intra pod within the DC.

In a certain way, yes

Kind Regards
-Lukas



On Apr 24, 2020, at 7:21 AM, Rabadan, Jorge (Nokia - US/Mountain View) <[e-mail]<mailto:[e-mail]>> wrote:

Hi Gyan,

The dci evpn overlay draft indeed provides that segmentation. EVPN routes are readvertised at the GWs with change in RD/VNI/Nhop, and this certainly optimizes the BUM replication. From end leaf nodes. The draft also introduces the use of an unknown Mac route that the GWs can advertise to their local POD, as opposed to readvertise all the received MAC routes. This can be used under the assumption that if a mac is unknown for a leaf, it must be somewhere beyond the GW. Finally, the draft also allows you to use an I-ES for multihoming and have all-active to two or more GWs.

Note that this draft has multiple implementations, and the only reason why is not an RFC yet is due to a normative reference that must be cleared first.

Thanks.
Jorge

From: Gyan Mishra <[e-mail]<mailto:[e-mail]>>
Date: Friday, April 24, 2020 at 3:54 PM
To: "Rabadan, Jorge (Nokia - US/Mountain View)" <[e-mail]<mailto:[e-mail]>>
Cc: BESS <[e-mail]<mailto:[e-mail]>>, Jeff Tantsura <[e-mail]<mailto:[e-mail]>>
Subject: Re: [bess] VXLAN BGP EVPN Question


Hi Jorge

I read through the draft and it sounds this vxlan segmentation is similar to multi site segmented multi part LSP used for DCI.   How  does this option compare or contrast with the multi site draft below.

With DCI evpn overlay you mentioned, the VNIs on the ASBRs are translated and not global.  Interesting.

With multi site the VNIs are Globally significant inter of intra site and an RT rewrite happens for the BGW to BGW middle segment to establish for the NVE to be stitched.

So with the DCI EVPN VNI translation does that end up netting the desired effect control plane segregation from data plane and providing that reduced size Mac VRF showing only active interesting traffic type 2 Mac-IP routes intra pod within the DC.

Multi site DCI
https://datatracker.ietf.org/doc/draft-sharma-multi-site-evpn/


Kind regards

Gyan

On Fri, Apr 24, 2020 at 3:07 AM Rabadan, Jorge (Nokia - US/Mountain View) <[e-mail]<mailto:[e-mail]>> wrote:
Hi Gyan,

If I may, note that:
https://tools.ietf.org/html/draft-ietf-bess-dci-evpn-overlay-10#section-4.6

Also provides vxlan segmentation, and while the description is based on DCI, you can perfectly use it for inter-pod connectivity.

Thanks.
Jorge

From: BESS <[e-mail]<mailto:[e-mail]>> on behalf of Gyan Mishra <[e-mail]<mailto:[e-mail]>>
Date: Friday, April 24, 2020 at 5:21 AM
To: Jeff Tantsura <[e-mail]<mailto:[e-mail]>>
Cc: BESS <[e-mail]<mailto:[e-mail]>>
Subject: Re: [bess] VXLAN BGP EVPN Question


Hi Jeff

Yes - Cisco has a draft for multi site for use cases capability of inter pod or inter site segmented path between desperate POD fabrics intra DC or as DCI option inter DC without MPLS.  The segmentation localizes BUM traffic and has border gateway DF election for BUM traffic that is segmented stitched between PODs as I mentioned similar to inter-as L3 vpn opt b.  There is a extra load as you said on the BGW border gateway performing the network vtep dencap from leaf and then again encap towards the egress border gateway.  Due to that extra load on the border gateway it’s not recommended to have spine function on BGW thus an extra layer for multi site to be scalable.  Definitely requires proprietary asic and not merchant silicon or white box solution.  The BUM traffic is much reduced as you stated from multi fabric connected super spine or single fabric spine that contains all leafs.  That decoupling sounds like incongruent control and data plane with Mac only Type 2 routes which would result in more BUM traffic  but it sounds like that maybe trade off of conversation learning only active flows versus entire data center wide Mac VRF being learned everywhere.  I wonder if their is an option to have that real decoupling of EVPN control plane and vxlan data plane overlay that does not impact convergence but adds stability and only active flow Type 2 Mac learner across the fabric.

https://datatracker.ietf.org/doc/draft-sharma-multi-site-evpn/

Kind regards

Gyan

On Thu, Apr 23, 2020 at 6:04 PM Jeff Tantsura <[e-mail]<mailto:[e-mail]>> wrote:
Gyan,

"Multi site” is not really an IETF terminology, this is a solution implement by NX-OS, there’s a draft though. Its main functionality is to localize VxLAN tunnels and provide segmented path vs end2end full mesh of VxLAN tunnels (participating in the same EVI). We are talking HER here.
The feature is heavily HW dependent as it requires BUM re-encapsulation at the boundaries (leaf1->BGW1-BGW2->leaf2..n). So good luck seeing it soon on low end silicon.
It doesn’t eliminate BUM traffic but significantly reduces the span of “broadcast domain” and reduces the need for large flood domains (modern HW gives you ~512 large flood groups, obviously depending on HW)

Wrt your question about Mac conversation learning - this is an implementation issue, nothing in EVPN specifications precludes you of doing so, moreover in the implementation I was designing (in my previous life) we indeed decoupled data plane learning from control plane advertisement so control plane was aware of “Active” flows.  Needless to say - this creates  an additional layer of complexity and all kinds of funky states in the system ;-).

Hope this helps

Cheers,
Jeff
On Apr 23, 2020, 8:30 AM -0700, Gyan Mishra <[e-mail]<mailto:[e-mail]>>, wrote:


Slight clarification with the arp traffic.  What I meant was broadcast traffic translated into BUM traffic with the EVPN architecture is there any way to reduce the amount of BUM traffic with a data center design requirement with vlan anywhere sprawl with 1000s of type 2 Mac mobility routes being learned between all the leaf VTEPs.

The elimination of broadcast is a tremendous gain and with broadcast suppression of multicast that does help but it would be nice to not have such massive Mac tables type 2 route churn chatter with a conversation learning where only active flows are are in the type 2 rib.

Kind regards

Gyan

On Wed, Apr 22, 2020 at 6:47 PM Gyan Mishra <[e-mail]<mailto:[e-mail]>> wrote:

In the description of the vxlan BGP evpn scenario has a typo on the multisite feature segmented LSP inter pod with the RT auto rewrite which is similar to MPLS inter-as option b not a.

Kind regards

Gyan

On Wed, Apr 22, 2020 at 5:57 PM Gyan Mishra <[e-mail]<mailto:[e-mail]>> wrote:

All

Had a question related to vxlan BGP EVPN architecture specifications defined in BGP EVPN NVO3 overlay RFC 8365 and VXLAN data plane RFC 7348.

In a Data Center environment where you have a multiple PODs individual fabrics per POD connected via a super spine extension using a Multi site feature doing auto rewrite of RTs to stitch the NVE tunnel between pods similar to inter-as option A.

So in this scenario where you have vlan sprawl everywhere with L2 and L3 VNIs everywhere as if it were a a single L2 domain.  The topology is a typical vxlan spine leaf topology where the L3 leafs are the TOR switch so very small physical  L2 fault domain. So I was wondering if with the vxlan architecture if this feature below is possible or if their is a way to do so in the current specification.

Cisco use to have a DC product called “fabric path” which was based on conversation learning.

Is there any way with existing vxlan BGP evpn specification or maybe future enhancement to have a Mac conversation learning capability so that only the active mac’s that are part of a conversations flow are the mac that are flooded throughout the vxlan fabric.  That would really help tremendously with arp storms so if new arp entries are generated locally on a leaf they are not flooded through the fabric unless their are active flows between leafs.

Also is there a way to filter type 2 Mac mobility routes between leaf switches at the control plane level based on remote vtep or maybe other parameters..  That would also reduce arp storms BUM traffic.

Kind regards

Gyan
--
Gyan  Mishra
Network Engineering & Technology
Verizon
Silver Spring, MD 20904
Phone: [phone]
Email: [e-mail]<mailto:[e-mail]>


--
Gyan  Mishra
Network Engineering & Technology
Verizon
Silver Spring, MD 20904
Phone: [phone]
Email: [e-mail]<mailto:[e-mail]>


--
Gyan  Mishra
Network Engineering & Technology
Verizon
Silver Spring, MD 20904
Phone: [phone]
Email: [e-mail]<mailto:[e-mail]>


_______________________________________________
BESS mailing list
[e-mail]<mailto:[e-mail]>
https://www.ietf.org/mailman/listinfo/bess
--
Gyan  Mishra
Network Engineering & Technology
Verizon
Silver Spring, MD 20904
Phone: [phone]
Email: [e-mail]<mailto:[e-mail]>


--
Gyan  Mishra
Network Engineering & Technology
Verizon
Silver Spring, MD 20904
Phone: [phone]
Email: [e-mail]<mailto:[e-mail]>


_______________________________________________
BESS mailing list
[e-mail]<mailto:[e-mail]>
https://www.ietf.org/mailman/listinfo/bess

--
Gyan  Mishra
Network Engineering & Technology
Verizon
Silver Spring, MD 20904
Phone: [phone]
Email: [e-mail]<mailto:[e-mail]>