Re: [Int-area] Comments on draft-ietf-intarea-frag-fragile-06

[Stewart Bryant <[e-mail]>]

On 29/01/2019 23:37, Fred Baker wrote:
>
>> Section 4.5:
>> "IP fragmentation causes problems for some routers that support Equal
>> Cost Multipath (ECMP). Many routers that support ECMP execute the
>> algorithm described in Section 4.4 in order to perform flow based
>> forwarding;
> As far as I know, routers that hash fields in the IP header to select a en ECMP next hop do so because all packets in a flow will hash the same way (modulo the issues with the transport port number), not because they are doing per-flow forwarding. The do so explicitly to avoid having to maintain per-flow state and yet make all fragments of a message follow the same path.

I agree with Fred. ECMP is normally done to distribute the load over the 
available next hops on a best effort basis. Originally it was done per 
packet, but that gave problems with out of order packet delivery, so the 
routers moved to doing it based on the five tuple described in this 
draft. It is a stateless best effort ECMP process with no regard to 
specific flows and the path for any five tuple may move arbitrarily if 
routing changes its mind on the ECMP set.

Fragmented packets are really bad news in networks that need ECMP. There 
is not enough entropy in the SA/DA/Protocol triplet and anything else 
results in misorder. But if ECMP is not done this overloads the default 
path.

MPLS is also stateless but there are more options, although the most 
common is to look past BoS to the five tuple, however some "features" 
make mistakes and look at a non-existent five tuple despite hints in the 
packet that thus is a bad idea.

>> therefore, the exhibit they same problematic behaviors
>> described in Section 4.4. In IPv6, the flow label may alternatively
>> used as input to the algorithm as opposed to parsing the transport
>> layer of packets to discern port numbers. The flow label should be
>> consistently set for a packets of flow including fragments, such that
>> a device does not need to parse packets beyond the IP header for the
>> purposes of ECMP."
>>
>> Add to section 7.3:
>>
>> "Routers SHOULD use IPv6 flow label for ECMP routing as described in [RFC6438]."

If we want to migrate to the FL then we really need to state that the FL 
MUST be set by the sender. Without, that we are never going to wean 
routers off looking at the five tuple, if indeed we ever succeed in 
doing that.

It we really need to fragment a packet, it would be better to stick the 
fragments inside a common UDP/IP(no frag) shim. Then the forwarders 
could carry on just as they are. We would never get misorder and we 
would not be faced with the impossible problem of changing the Internet 
core forwarding behaviour to a single consistent model.

- Stewart

>>
>> _______________________________________________
>> Int-area mailing list
>> [e-mail]
>> https://www.ietf.org/mailman/listinfo/int-area
> --------------------------------------------------------------------------------
> Victorious warriors win first and then go to war,
> Defeated warriors go to war first and then seek to win.
>       Sun Tzu
>
>
> _______________________________________________
> Int-area mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/int-area