Re: [Lsr] New draft on Flex-Algorithm Bandwidth Constraints

[Robert Raszuk <[e-mail]>]

Peter,

Completely disagree.

Real say enterprise networks are being build with emulated circuits
(example VPWS). In one company I was working at it was about 80% of
emulated links all over the world in their WAN. Yes for me it was a shock
as I did not realize how much this emulated links took over the world. In
most geographies you even can not get any link of less then 10Gbps to be
real any more. Only emulated option is on the table.

Emulated circuits run over someone's IP backbones. You can understand the
consequences of this. Not only link delay changes a lot but you run into
very interesting set of issues.

Maybe you think of the backbones of the 90s or 2000s where dark fiber or
SDH or SONET were in use for interconnects.

Well no more. IETF came with such brilliant ideas to emulate L2 over L3 and
here we go.

Reality is not what we wish it to be.

Cheers,
R.


On Thu, Mar 4, 2021 at 10:07 AM Peter Psenak <[e-mail]> wrote:

> Robert,
>
> On 03/03/2021 20:57, Robert Raszuk wrote:
> > Peter,
> >
> >  >  that differ by few microsecond
> >
> > Really you normalize only single digit microseconds ???
> >
> > What if link delay changes in milliseconds scale ? Do you want to
> > compute new topology every few milliseconds ?
>
> let me repeat again.
>
> Min delay is not something that changes every few milliseconds
> significantly. It's a semi static variable that reflects the property of
> the underlying physical path. It only changes when the physical path
> properties changes - e.g. the optical path reroutes, etc. We
> deliberately picked Min delay for flex-algo purposes for this semi
> static property.
>
> The small, non significant changes can be filtered by the normalization.
>
> If the min delay changes significantly every few milliseconds there's
> something wrong with the link itself - we have standard dampening
> mechanisms in LS protocols to deal with unstable links that would kick
> in. Similar to what we do if the link flaps every few milliseconds.
>
> >
> > Out of curiosity as this is not a secret -  What are your default min
> > delay normalization timers (if user does not overwrite with their own).
>
> there is no timer needed for the normalization itself.
>
> You are likely referring TWAMP computation interval which is 30 sec,
> with probes being sent every 3 seconds in our implementation by default,
> if I'm not mistaken.
>
> Normalization is applied to the value that come from the above.
>
> thanks,
> Peter
>
>
>
> > Likewise how Junos or Arista normalizes it today ?
> >
> > Thx,
> > R.
> >
> > On Wed, Mar 3, 2021 at 7:41 PM Peter Psenak <[e-mail]
> > <mailto:[e-mail]>> wrote:
> >
> >     Tony,
> >
> >     On 03/03/2021 19:14, Tony Li wrote:
> >      >
> >      > Peter,
> >      >
> >      >>> There are several link types in use that exhibit variable
> >     delay: satellite links (e.g., Starlink), microwave links, and
> >     ancient link layers that deliver reliability through retransmission.
> >      >>> Any of these (and probably a lot more) can create a noticeable
> >     and measurable difference in TWAMP. That would be reflected in an FA
> >     metric change. If you imagine a situation with multiiple parallel
> >     paths with nearly identical delays, you can easily imagine an
> >     oscillatory scenario.   IMHO, this is an outstanding concern with FA.
> >      >> yes, and that is what I referred to as "delay normalization",
> >     which can avoid that oscillation.
> >      >
> >      >
> >      > It can also negate the benefits of the feature. One might well
> >     imagine that Starlink would want to follow a min-delay path for
> >     optimality.  If the delay variations are “normalized” out of
> >     existence, then the benefits are lost.  The whole point is to track
> >     the dynamics.
> >
> >     for all practical purposes that we use it for, the two values of min
> >     delay that differ by few microsecond can be treated as same without
> any
> >     loss of functionality. So it's about the required normalization
> >     interval
> >     - something that can be controlled by the user.
> >
> >     thanks,
> >     Peter
> >
> >
> >
> >      >
> >      >
> >      >>> Please note that I’m NOT recommending that we back away.
> >     Rather, we should seek to solve the long-standing issue of
> >     oscillatory routing.
> >      >>
> >      >> not that I disagree. History tells us that the generic case of
> >     oscillation which is caused by the traffic itself is a hard problem
> >     to solve.
> >      >
> >      >
> >      > Any oscillation is difficult to solve.  Positive feedback
> >     certainly can exacerbate the problem. But solving hard problems is
> >     why we are here.
> >      >
> >      > Yours in control theory,
> >      > Tony
> >      >
> >      >
> >      >
> >
>
>