Re: [tcpm] [v6ops] Flow Label Load Balancing

[Tom Herbert <[e-mail]>]

Hi Fernando, comments in line...

On Wed, Nov 25, 2020 at 12:13 AM Fernando Gont <[e-mail]> wrote:
>
> Hi, Tom,
>
> On 24/11/20 16:43, Tom Herbert wrote:
> [....]
> > Modulating the flow label is a means to affect the routing of packets
> > through the network that uses flow labels as input to the ECMP hash.
>
> What's the point?
>
> 1) You cannot tell *if* the FL is being used.
>
Generally true, but in a limited domain this information could be
discerned. I'd note that it's also generally true that we don't know
if there is a load balancer or stateful firewall in the path that
requires consistent routing, but in a limited domain we could know
that also.

> 2) Changing the FL does not necessarily mean that packets will employ a
> different link.

It's an opportunistic mechanism. If a connection is failing and we get
a better path that fixes it by simply changing the flow label then
what's the harm?

>
> 3) If the network is failing, shouldn't you handle this via routing?
>
Sure, but then that requires an out of band feedback loop from a TCP
implementation to the network infrastructure to indicate there is a
problem and then the network needs to respond. That's significant
infrastructure and higher reaction time than doing something in TCP
and IP. Think of modulating the flow label is an inexpensive form of
source routing within a limited domain that doesn't need any
infrastructure or heavyweight protocols or something like segment
routing.

>
>
> > The basic idea is that the flow label associated with a connection is
> > randomly changed when the stack observes that the connection is
> > failing (e.g. and an RTO). There is nothing in the specs that prevents
> > this since the source is at liberty to set the flow label as it sees
> > fit.
>
> The FL is expected to remain constant for the life of a flow. A
> retransmitted packet is part of the same flow as the
> originally-transmitted packet. So this seems to be contradicting the
> very specification of the FL.
>
> For instance, If a RTO for a flow causes the FL to change, then one may
> possibly argue that the FL is not naming/labeling what is said/expected
> to be anming/labeling.

Specifically, RFC6437 states:

"It is therefore RECOMMENDED that source hosts support the flow label
by setting the flow label field for all packets of a given flow to the
same value chosen from an approximation to a discrete uniform
distribution."

So that is clearly a just recommendation, and not a requirement (and
definitely not a MUST). Furthermore, RFC6437 states:

"A forwarding node MUST either leave a non-zero flow label value
unchanged or change it only for compelling operational security
reasons as described in Section 6.1."

So there's no guarantee in the protocol specs that flow labels are
consistent for the life of the connection, which means that the
network cannot assume that and thus it would be incorrect if the
network tried to enforce flow label consistency as a protocol
requirement. As I said, it is prudent to try to be consistent with
flow labels and the default behavior in Linux should be changed,
however I do not believe there's a valid claim of non-conformance that
motivates removal of the feature that is already deployed.

Tom




>
>
>
> > The feature is useful in large datacenter networks, like
> > pparently Facebook where the patches originate, since information
> > discerned by TCP can opportunistically be applied to route selection.
> > The practical issue is that there are stateful devices like firewalls
> > that require consistent routing in the network in which case changing
> > the flow label can confuse them. As I mentioned, the original intent
> > was that the flow label randomization feature should be opt-in instead
> > of on by default.
>
> So... where is the "source" of the packet that would be "modulating" the FL?
>
> Thanks,
> --
> Fernando Gont
> e-mail: [e-mail] || [e-mail]
> PGP Fingerprint: 7809 84F5 322E 45C7 F1C9 3945 96EE A9EF D076 FFF1
>
>
>