Re: [bess] draft-hao-bess-inter-nvo3-vpn-optionc

[Haoweiguo <[e-mail]>]

Hi Wim,
Pls see inline with [weiguo2].
weiguo
________________________________________
From: BESS [[e-mail]] on behalf of Henderickx, Wim (Wim) [[e-mail]]
Sent: Thursday, November 19, 2015 23:02
To: Thomas Morin; [e-mail]
Subject: Re: [bess] draft-hao-bess-inter-nvo3-vpn-optionc
Thomas, thx for the time summarising this and great summary.


On 19/11/15 02:52, "BESS on behalf of Thomas Morin" <[e-mail] on behalf of [e-mail]> wrote:
>I don't believe the discussion can be usefully summarized in terms of
>"fatal flaws".
>Let me try to summarize my understanding of the thing discussed in this
>thread.
>
>The solution in the draft has an impact on ASBR hardware due to the new
>kind of stitching required (as well summarized by Diego). Whether this
>is a big issue or not depends on the vendor. The solution in the draft,
>for the NVE part, would be implementable in software quite easily. But,
>because of the limitations in the widespread ToR chipsets for VXLAN (and
>their lack of support for MPLS/(GRE or UDP)), its multiple-UDP-port
>variant would be harder to implement in ToRs or at least not without a
>performance hit (people who know that better, feel free to correct).
WH> some HW cannot do this at all, so lets dismiss this idea.
[weiguo2]:  Currently TOR switches can't realize multiple-UDP-port variant VXLAN encapsulation, but we can't give conclusion that hardware can't realize this function forever.
The TOR only need to do the a little more complicated encap work for the traffic from DC to WAN direction. For the traffic from WAN to DC direction, the encapsulation is normal VXLAN encapsulation, it has no extra decap requirements for the TOR switches. Normally encapsulation process is simpler than decapsulation. The more complicated decap work is performed at ASBR-d which is router device and has more flexibility.
Currently MPLS+MPLS+GRE encap and decap process may be possible to be realized relying on internal loop in TOR switch, but just as i have mentioned before, many switch doesn't use MPLS over GRE for north-south bound traffic forwarding, they just want to use VXLAN encapsulation for both north-south and east-west bound traffic. I think variant UDP port solution is reasonable, it can't be dismissed.

>The multiple-loopback approach has operational drawbacks, but whether or
>not these are killer issues may depend on the targeted scale (in terms
>of number of NVEs) and may depend on other factors (operator practices,
>ability to automate ASBR configs).
WH> ok now we have not discussed the constraints some HW vendors have with respect to global VNIDs. To make this work all VNID/Labels need to be globally unique. Hmmmmm
[weiguo2]:  In SDN scenario, a virtual network normally is represented by a global VN ID or MPLS VPN Label to simplify network management. I think it's not strange to allocate global unique MPLS VPN Label or VN ID for a virtual network now.
>
>The alternative approach, put forward by Wim, consist in using existing
>Option-C with a variant of the 3-label variant relying on an
>MPLS/MPLS/UDP-or-GRE encap (instead of the 3-label MPLS stack).  This
>approach would not necessitate new standardized procedures, would
>require less changes on ASBRs. However it is not supported today by
>vswitches or ToRs (unless the bottom MPLS label is passed to the VM,
>which I think would restrict the application to a subset of the
>scenarios for which Option C is interesting). Having it supported in
>vswitches may be a simple matter of writing the software, but ToR
>support seems to require evolution in chipsets ; whether such a support
>is likely to appear hasn't been discussed in the thread.
>
>All in all, it seems there is no solution to cover an Option C scenario
>with today's generation ToRs, and the question for tomorrow's generation
>ToRs hasn't been really discussed.
>
>Based on the above, I would think the question boils down to whether it
>is desirable to specify an Option C variant usable with vswitches as
>they are today and possibly usable with a performance hit with today's
>ToRs chipsets, at the expense of a required evolution on ASBRs.  The
>alternative requiring some evolution on vswitches and waiting for future
>ToRs chipsets.
WH> I vote for a an evolution of switches/TORs that have proper support for this.
I hope some HW vendors of TOR chips shime in, but I am told the MPLS solution is possible in the next generation chips they are working on.
[weiguo2]:  I think it's better not to change current TOR hardware to realize option-C interconnection. No TOR hardware modification solution must be provided. But for future case, we can propose extra hardware requirements for TORs.
The other options requires baggage forever on the ASBR elements that does not bring value and it is better to avoid this and build an architecture which is future proof and more cost effective. My 2 cents
>
>-Thomas

>
>Zhuangshunwan  :
>>
>> Hi Diego,
>>
>> Thanks for your comments. Pls see inline with [Vincent].
>>
>> Vincent
>>
>> *发件人:*BESS [mailto:[e-mail]] *代表 *Diego Garcia del Rio
>> *发送时间:*2015年11月18日14:25
>> *收件人:*[e-mail]
>> *主题:*Re: [bess] draft-hao-bess-inter-nvo3-vpn-optionc
>>
>> Some comments from my side,
>> I think the draft makes quite a few assumptions on specific
>> implementation details that are way too general to be considered
>> widely available.
>> Most of the TOR devices today already pay a double-pass penalty when
>> doing routing of traffic in/out of vxlan-type tunnels. Only the newest
>> generation can route into tunnels without additional passes. And are
>> definitively limited in being able to arbitrary select UDP ports on a
>> per tunnel basis. In fact, most are even limited at using more than
>> one VNID per "service" of sorts.
>> [Vincent]: Yes, the new generation BCM chipset has already supported
>> VXLAN routing without additional passes. For OVS/TOR, VXLAN
>> encapsulation is more popular than MPLSoGRE/UDP, and has better
>> performance.
>> The IP-addressed based implementation which would, I assume, imply
>> assigning one or more CIDRs to a loopback interface on the ASBR-d is
>> also quite arbitrary and does not seem like a technically "clean"
>> solution. (besides burning tons of IPs). As a side-note, most PE-grade
>> routers i've worked with were quite limited in terms of IP addresses
>> used for tunnel termination and it wasn't that just a simple pool can
>> be used.
>> [Vincent]: I think the larger VTEP IP address range on ASBR-d has no
>> limitations. For the hardware on ASBR-d, it has capability to
>> terminate multiple VXLAN tunnels with arbitrary destination VTEP IP
>> addresses.
>> Wim's mentions on cases where the Application itself, hosted in a
>> datacenter, would be part of the option-C interconnect, was dismissed
>> without much discussion so far, while, if we look in detail at the
>> type of users which will even consider a complex topology like model-C
>> its most likely users and operators very familiar with MPLS VPNs in
>> the WAN. Those type of operators will most likely be very interested
>> in deploying MPLS or WAN-grade applications (i.e., virtual-routers or
>> other VNFs) in the DC and thus its highly likely that the interconnect
>> would not terminate at the NVE itself but rather the TS (the virtual
>> machine).
>> Also, the use of UDP ports at random would imply quite complex logic
>> on the ASBR-d IMHO. Im not saying its impossible, but since the
>> received packet now not only has to be received on a random (though
>> locally chosen) UDP port and this information preserved in the
>> pipeline to be able to do the double-tunnel-stitching, it sounds quite
>> complex. I am sure someone in the list will mention this has already
>> been implemented somewhere, and I won't argue with that. And let's not
>> even bring into account what this would do to any DC middlebox that
>> now has to look at vxlan over almost any random port. We have to go
>> back to the "is it a 4 or is it a 6 in byte x" heuristics to try to
>> guess whether the packet is vxlan or just something entirely different
>> running over IP.
>> [Vincent]: Using NP or multi-core CPU hardware technology, it can be
>> implemented although deeper packet inspection is needed to perform UDP
>> port and MPLS stitching.
>> In general I feel the proposed solution seems to be fitting of a
>> specific use-case which is not really detailed in the draft and does
>> not describe   a solution that would "elegantly" solve the issues at
>> hand. It just feels like we're using any available bit-space to stuff
>> data into places that do not necesarily belong.
>> Yes, MPLS encapsulations on virtual switches are not yet fully
>> available, and there can be some performance penalty on the TORs, but
>> the solutions are much cleaner from a control and data plane point of
>> view. Maybe I'm too utopic.
>> [Vincent]: I think pure VXLAN solution has its scenario, it's general
>> rather than specific. We can't require all OVS/NVEs support VXLAN +
>> MPLSoGRE at the same time.
>> Best regards,
>> Diego
>> ---------------------------------------------------------------------------------
>> Hi,
>> The problem we are trying to solve is to reduce data center
>> GW/ASBR-d's forwarding table size, the motivation is same as
>> traditional MPLS VPN option-C. Currently, the most common practise on
>> ASBR-d is to terminate VXLAN encapsulation, look up local routing
>> table, and then perform MPLS encapsulation to the WAN network. ASBR-d
>> needs to maintain all VM's MAC/IP. In Option-C method, only transport
>> layer information needed to be maintained at GW/ASBR-d, the
>> scalability will be greatly enhanced. Traditonal Option-C is only for
>> MPLS VPN network interworking, because VXLAN is becoming pervasive in
>> data center,  the solution in this draft was proposed for the
>> heterogeneous network interworking.
>> The advantage of this solution is that only VXLAN encapsulation is
>> required for OVS/TOR. Unlike Wim's solution, east-west bound traffic
>> uses VXLAN encap, while north-south bound traffic uses MPLSoGRE/UDP encap.
>> There are two solutions in this draft:
>> 1. Using VXLAN tunnel destination IP for stitching at ASBR-d.
>> No data plane modification requirements on OVS or TOR switches, only
>> hardware changes on ASBR-d. ASBR-d normally is router, it has
>> capability to realize the hardware changes. It will consume many IP
>> addresses and the IP pool for allocation needs to be configured on
>> ASBR-d beforehand.
>> 2. Using VXLAN destination UDP port for stitching at ASBR-d.
>> Compared with solution 1, less IP address will be consumed for
>> allocation. If UDP port range is too large, we can combine with
>> solution 1 and 2.
>> In this solution, both data plane modification changes are needed at
>> OVS/TOR and ASBR-d. ASBR-d also has capability to realize the hardware
>> changes. For OVS, it also can realize the data plane changes. For TOR
>> switch, it normally can't realize this function.  This solution mainly
>> focuses on pure software based overlay network, it has more
>> scalability. In public cloud data center, software based overlay
>> network is the majority case.
>> Whether using solution 1 or 2 depends on the operators real envionment.
>> So I think our solution has no flaws, it works fine.
>> Thanks,
>> weiguo
>> ________________________________
>> From: BESS [[e-mail] <mailto:[e-mail]>] on
>> behalf of John E Drake [[e-mail] <mailto:[e-mail]>]
>> Sent: Wednesday, November 18, 2015 2:49
>> To: Henderickx, Wim (Wim); EXT - [e-mail]
>> <mailto:[e-mail]>; BESS
>> Subject: Re: [bess] draft-hao-bess-inter-nvo3-vpn-optionc
>> Hi,
>> I think Wim has conclusively demonstrated that this draft has fatal
>> flaws and I don’t support it.  I also agree with his suggestion that
>> we first figure out what problem we are trying to solve before solving it.
>> Yours Irrespectively,
>> John
>> From: BESS [mailto:[e-mail]
>> <mailto:[e-mail]>] On Behalf Of Henderickx, Wim (Wim)
>> Sent: Tuesday, November 17, 2015 12:49 PM
>> To: EXT - [e-mail] <mailto:[e-mail]>; BESS
>> Subject: Re: [bess] draft-hao-bess-inter-nvo3-vpn-optionc
>> ― Snip ―
>> No, the spec as it is can be implemented in its VXLAN variant with
>> existing vswitches (e.g. OVS allows to choose the VXLAN destination
>> port, ditto for the linux kernel stack).
>> (ToR is certainly another story, most of them not having a flexible
>> enough VXLAN dataplane nor support for any MPLS-over-IP.)
>> WH> and how many ports simultaneously would they support? For this to
>> work every tenant needs a different VXLAN UDP destination port/receive
>> port.
>> There might be SW elements that could do some of this, but IETF
>> defines solutions which should be implemented across the board
>> HW/SW/etc. Even if some SW switches can do this, the proposal will
>> impose so many issues in HW/data-plane engines that I cannot be behind
>> this solution.
>> To make this work generically we will have to make changes anyhow.
>> Given this, we better do it in the right way and guide the industry to
>> a solution which does not imply those complexities. Otherwise we will
>> stick with these specials forever with all consequences (bugs, etc).
>> - snip -
>> From: "[e-mail]
>> <mailto:[e-mail]><mailto:[e-mail]
>> <mailto:[e-mail]>>" <[e-mail]
>> <mailto:[e-mail]><mailto:[e-mail]
>> <mailto:[e-mail]>>>
>> Organization: Orange
>> Date: Tuesday 17 November 2015 at 01:37
>> To: Wim Henderickx <[e-mail]
>> <mailto:[e-mail]><mailto:[e-mail]
>> <mailto:[e-mail]>>>, BESS <[e-mail]
>> <mailto:[e-mail]><mailto:[e-mail] <mailto:[e-mail]>>>
>> Subject: Re: [bess] draft-hao-bess-inter-nvo3-vpn-optionc
>> Hi Wim, WG,
>> 2015-11-16, Henderickx, Wim (Wim):
>> 2015-11-13, Henderickx, Wim (Wim):
>> Thomas, we can discuss forever and someone need to describe
>> requirements, but the current proposal I cannot agree to for the
>> reasons explained.
>> TM> Well, although discussing forever is certainly not the goal, the
>> reasons for rejecting a proposal need to be thoroughly understood.
>> WH> my point is what is the real driver for supporting a plain VXLAN
>> data-plane here, the use cases I have seen in this txt is always where
>> an application behind a NVE/TOR is demanding option c, but none of the
>> NVE/TOR elements.
>> My understanding is that the applications  are contexts where overlays
>> are present is when workloads (VMs or baremetal) need to be
>> interconnected with VPNs. In these contexts, there can be reasons to
>> want Option C to reduce the state on ASBRs.
>> In these context, its not the workload (VM or baremetal) that would
>> typically handle VRFs, but really the vswitch or ToR.
>> WH2> can it not be all cases: TOR/vswitch/Application. I would make
>> the solution flexible to support all of these not?
>> 2015-11-13, Henderickx, Wim (Wim):
>> TM> The right trade-off to make may in fact depend on whether you prefer:
>> (a) a new dataplane stitching behavior on DC ASBRs (the behavior
>> specified in this draft)
>> or (b) an evolution of the encaps on the vswitches and ToRs to support
>> MPLS/MPLS/(UDP or GRE)
>> WH> b depends on the use case
>> I don't get what you mean by "b depends on the use case".
>> WH> see my above comment. If the real use case is an application
>> behind NVE/TOR requiring model C, than all the discussion on impact on
>> NVE/TOR is void. As such I want to have a discussion on the real
>> driver/requirement for option c interworking with an IP based Fabric.
>> Although I can agree than detailing requirements can always help, I
>> don't think one can assume a certain application to dismiss the proposal.
>> WH> for me the proposal is not acceptable for the reasons explained:
>> too much impact on the data-planes
>> I wrote the above based on the idea that the encap used in
>> MPLS/MPLS/(UDP or GRE), which hence has to be supported on the ToRs
>> and vswitches.
>> Another possibility would be service-label/middle-label/Ethernet
>> assuming an L2 adjacency between vswitches/ToRs and ASBRs, but this
>> certainly does not match your typical DC architecture. Or perhaps had
>> you something else in mind ?
>> WH> see above. The draft right now also requires changes in existing
>> TOR/NVE so for me all this discussion/debate is void.
>> No, the spec as it is can be implemented in its VXLAN variant with
>> existing vswitches (e.g. OVS allows to choose the VXLAN destination
>> port, ditto for the linux kernel stack).
>> (ToR is certainly another story, most of them not having a flexible
>> enough VXLAN dataplane nor support for any MPLS-over-IP.)
>> WH> and how many ports simultaneously would they support?
>> WH> and depending on implementation you don’t need to change any of
>> the TOR/vswitches.
>> Does this mean that for some implementations you may not need to
>> change any of the TOR/vswitches, but that for some others you may ?
>> WH> any proposal on the table requires changes, so for me this is not
>> a valid discussion
>> See above, the proposal in the draft does not necessarily need changes
>> in vswitches.
>> Let me take a practical example : while I can quite easily see how to
>> implement the procedures in draft-hao-bess-inter-nvo3-vpn-optionc
>> based on current vswitch implementations of VXLAN, the lack of
>> MPLS/MPLS/(UDP, GRE) support in commonplace vswitches seems to me as
>> making that alternate solution you suggest harder to implement.
>> WH> I would disagree to this. Tell me which switch/TOR handles
>> multiple UDP ports for VXLAN ?
>> I mentioned _v_switches, and many do support a variable destination
>> port for VXLAN, which is sufficient to implement what the draft proposes.
>> -Thomas
>> From: Thomas Morin <[e-mail]
>> <mailto:[e-mail]><mailto:[e-mail]
>> <mailto:[e-mail]>>>
>> Organization: Orange
>> Date: Friday 13 November 2015 at 09:57
>> To: Wim Henderickx <[e-mail]
>> <mailto:[e-mail]><mailto:[e-mail]
>> <mailto:[e-mail]>>>
>> Cc: "[e-mail] <mailto:[e-mail]><mailto:[e-mail]
>> <mailto:[e-mail]>>" <[e-mail]
>> <mailto:[e-mail]><mailto:[e-mail] <mailto:[e-mail]>>>
>> Subject: Re: [bess] draft-hao-bess-inter-nvo3-vpn-optionc
>> Hi Wim,
>> I agree on the analysis that this proposal is restricted to
>> implementations that supports the chosen encap with non-IANA ports
>> (which may be hard to achieve for instance on hardware
>> implementations, as you suggest), or to context where managing
>> multiple IPs would be operationally viable.
>> However, it does not seem obvious to me how the alternative you
>> propose [relying on 3-label option C with an MPLS/MPLS/(UDP|GRE)
>> encap] addresses the issue of whether the encap behavior is supported
>> or not (e.g. your typical ToR chipset possibly may not support this
>> kind of encap,  and even software-based switches may not be ready to
>> support that today).
>> My take is that having different options to adapt to various
>> implementations constraints we may have would have value.
>> (+ one question below on VXLAN...)
>> -Thomas
>> 2015-11-12, Henderickx, Wim (Wim):
>> On VXLAN/NVGRE, do you challenge the fact that they would be used with
>> a dummy MAC address that would be replaced by the right MAC by a
>> sender based on an ARP request when needed ?
>> Is the above the issue you had in mind about VXLAN and NVGRE ?
>> WH> yes
>> I you don't mind me asking : why do you challenge that ?
>>
>
>_______________________________________________
>BESS mailing list
>[e-mail]
>https://www.ietf.org/mailman/listinfo/bess
_______________________________________________
BESS mailing list
[e-mail]
https://www.ietf.org/mailman/listinfo/bess