Hi all,
I am writing to this alias because I do not know the proper one for such type of a problem (OS/LINUX/BSD).
The history of how Alexander Azimov (Yandex) has found the problem is below.
In short: if TCP loses connectivity for 200ms (or 1s in SYN stage) then TCP changes IPv6 flow label (for the active TCP session!) to push traffic to a different path.
Current networks are extensively ECMP, if intermediate nodes support flow label for hash calculation then a high probability that the path would be changed.
LINUX/BSD does not want to wait till the network will fix its problem.
If the final destination was anycast then the final destination would be changed too by the same hash calculation.
The stateful session would be broken as a result (see the second part of Alexander’s presentation below).
Since the time LINUX has made the default RTO flow label recalculation (2016), IPv6 Anycast is broken.
People would have one more reason not to migrate to IPv6. Flow label does not exist in IPv4 – OS is not capable to break IPv4 Anycast similarly.
Is anybody would like to spend his/her karma to save IPv6 Anycast OR let it die?
It is broken already for 5 years and nobody has spotted it up to now. Is it needed?
(I have seen a few drafts heavily dependent on IPv6 anycast)
What is proper WG for such a problem?
I am concerned that Anycast has been killed, it is not an easily replaceable tool.
Maybe somebody would propose something better but if not
then LINUX should be returned to 2015 when flow label change on RTO was a non-default configuration.
Such LINUX behavior could be valuable in some restricted domains (see below) when the administrator is sure that Anycast is not possible on the traffic path.
Eduard
From: Vasilenko Eduard
Sent: Tuesday, August 3, 2021 12:05 PM
To: 'Jeff Tantsura' <jefftant.ietf@xxxxxxxxx>; Alexander Azimov <a.e.azimov@xxxxxxxxx>
Cc: Alexander Azimov <mitradir@xxxxxxxxxxxxxx>; routing WG <rtgwg@xxxxxxxx>
Subject: RE: Self-healing Networking with Flow Label
Hi all,
Not many people worldwide read this alias and understand
That RTO could be leveraged to fight “silent drops” in the DC environment.
It is a good use case to publish/document (with more details that it was in the presentation).
I hope that in the future OAM would be used for this purpose – it is better from architecture point of view.
Eduard
From: Jeff Tantsura [mailto:jefftant.ietf@xxxxxxxxx]
Sent: Tuesday, August 3, 2021 1:08 AM
To: Alexander Azimov <a.e.azimov@xxxxxxxxx>
Cc: Vasilenko Eduard <vasilenko.eduard@xxxxxxxxxx>; Alexander Azimov <mitradir@xxxxxxxxxxxxxx>; routing WG <rtgwg@xxxxxxxx>
Subject: Re: Self-healing Networking with Flow Label
Eduard,
The idea of the draft to come is to explain what to do - when and how.
The goal is not to regulate (we really don’t) but to provide, similarly to RFC7938 a set of guidelines that community can use to build better and more resilient networks.
Cheers,
Jeff
On Aug 2, 2021, at 04:01, Alexander Azimov <a.e.azimov@xxxxxxxxx> wrote:
Eduard,
пн, 2 авг. 2021 г. в 13:45, Vasilenko Eduard <vasilenko.eduard@xxxxxxxxxx>:
It is the key in this presentation “This behavior MUST be switched off by default”
It has been shown on slides 7-10 that flow label change on RTO is enabled by default for BSD and LINUX.
It needs regulation. It needs a standard RFC. Because it kills Anycast otherwise.
As I'm partially responsible for the key points of the presentation, I can stress that it is a bit different.
- We have an opportunity for self-healing TCP on top of IPv6, it should be preserved;
- The Linux defaults should be changed to a safe mode to prevent session timeouts;
- The hash recalculation behavior should be documented;
I'm not sure what you mean by the term 'regulation'.
The story of how to use RTO to work-around “silent drop” vendor’s bugs could be a good informational RFC.
My be people developing iOAM would pay more attention to this use case.
IMHO: these are 2 separate drafts.
I'm not sure about it, we'll try to provide -00 before the next IETF meeting, let's see how it progresses.
Eduard
From: Alexander Azimov [mailto:mitradir@xxxxxxxxxxxxxx]
Sent: Monday, August 2, 2021 1:20 PM
To: Vasilenko Eduard <vasilenko.eduard@xxxxxxxxxx>; Jeff Tantsura <jefftant.ietf@xxxxxxxxx>
Cc: routing WG <rtgwg@xxxxxxxx>
Subject: Re: Self-healing Networking with Flow Label
Eduard,
Please see the quote from the slide 28. My suggestion was:
Client – sends SYN, Server – responds with SYN&ACK
- In case of SYN_RTO or RTO events Server SHOULD recalculate its TCP socket hash, thus change Flow Label. This behavior MAY be switched on by default;
- In case of SYN_RTO or RTO events Client MAY recalculate its TCP socket hash, thus change Flow Label. This behavior MUST be switched off by default;
This looks like a safe default behavior, that saves the part of the improvements, but makes the work with stateful anycast services safe.
And yes, IMO it's ok to have a knob to enable it in the controlled environment. If you ask how to enable it in the presence of internal anycast services - there was also a suggestion in the slides: eBPF. It gives a good way to make this kind of separation.
02.08.2021, 11:48, "Vasilenko Eduard" <vasilenko.eduard@xxxxxxxxxx>:
Hi Jeff,
The situation when Control Plane does not understand what the Forwarding pane doing is a bug.
Yes, RTO in TCP helps to find a work-around for this bug. And yes, Anycast is typically absent inside DC – it does not create the problem in the DC environment.
But the same LINUX is used outside DC. RTO Flow Label change here would create even more problems if Anycast would happen on the traffic path (not much predictable for client).
Do we need separate LINUX distribution for DC and separate distribution for other environments?
Or should we rely on the proper non-default configuration for different environments? (Admin should not forget to change)
What if Anycast would become needed in DC?
RTO flow label recalculation is mutually exclusive with Anycast on the traffic part.
What is more valuable for the public?
IMHO: It is better to fight the problem of such type of a bug with iOAM than to cancel Anycast.
IMHO: It is better to have Flow Label recalculation on RTO as “off” by default.
DC Admin has the higher qualification to activate it in a controlled environment than every client worldwide that should not forget to disable it.
Eduard
From: Jeff Tantsura [mailto:jefftant.ietf@xxxxxxxxx]
Sent: Monday, August 2, 2021 6:56 AM
To: Vasilenko Eduard <vasilenko.eduard@xxxxxxxxxx>
Cc: mitradir@xxxxxxxxxxxxxx; routing WG <rtgwg@xxxxxxxx>
Subject: Re: Self-healing Networking with Flow Label
Eduard,
The issue is present not in link/device case, if well implemented - fast rehash takes care of updating forwarding within a number of ms. The problem is with “gray” failures, when the link in question is UP from routing/forwarding prospective but drops traffic (mostly occasionally and when a HW bug occurs has a distinct flow attributes).
In many large DC fabrics, the majority of the traffic is east-west, between end-points that aren’t anycast. In such deployments - the solution solves issues rather elegantly and without any interventions from the operator.
The issues/side effects are well understood and will be documented.
The best way to receive RTGWG emails is well… to subscribe to RTGWG ;-)
Cheers,
Jeff
On Aug 1, 2021, at 09:47, Vasilenko Eduard <vasilenko.eduard@xxxxxxxxxx> wrote:
Hi Alexander,
Have I understood your presentation right?
The client SHOULD change IPv6 flow label after SYN RTO to have a chance to be moved to the working path inside DC fabric (if DC fabric supports flow label for hash calculation)
But at the same time
The client SHOULD NOT change the IPv6 flow label after SYN RTO to avoid being switched to a different TCP proxy engine.
Looks like a deadlock, especially if both things should happen for the same traffic:
it should reach DC fabric
and
it should be hash load-balanced between different TCP proxy engines (or applications) inside DC Fabric.
I see one bad solution (“Disable Flow Label”):
Routers up to TCP proxy engine SHOULD be configured not to use flow label (by the way these are all routers on the Internet),
TCP flow engines SHOULD be outside of the DC Fabric (CLOS) – probably in front of it.
Routers/Switches inside DC Fabric SHOULD use flow labels.
I see another bad solution (“Disable Anycast”):
Disable anycast on routers in principle, use only stateful LB.
It has been commented in the chat that Anycast is not possible in principle for stateful connection. It is too general a statement.
Anycast is just not compatible with Flow Label. It is not a problem for IPv4 anycast even if the connection is stateful (TCP) because 5-tuple for hash would not change.
Hence, IPv6 anycast has become dead at the time when Flow Label change has been added in LINUX for active TCP session.
Among 3 thins:
- Anycast
- Flow Label load balancing (basic Flow Label functionality)
- Flow Label change on the active session for application to be more active in new path search
You have to choose which one to kill – all 3 are not compatible with each other at the same.
I vote to disable Flow Label change in LINUX. Then wait till the network would fix itself.
We have so many fancy TE tools our days. A broken link or a broken node could be excluded from routing for 50ms.
PS: I am not subscribed to the RTGWG alias, please keep me on a copy of this thread.
<image001.png>
Best Regards
Eduard Vasilenko
Senior Architect
Europe Standardization & Industry Development Department
Tel: +7(985) 910-1105, +7(916) 800-5506
_______________________________________________
rtgwg mailing list
rtgwg@xxxxxxxx
https://www.ietf.org/mailman/listinfo/rtgwg
--
Best regards,
Alexander Azimov
_______________________________________________
rtgwg mailing list
rtgwg@xxxxxxxx
https://www.ietf.org/mailman/listinfo/rtgwg
--
Best regards,
Alexander Azimov
Does ANYCAST UDP work?
It would not make me in the least bit sad if ANYCAST TCP doesn't work because I don't think it should.
ANYCAST for DNS and other naming/discovery schemes makes perfect sense. If your discovery scheme requires more than one request packet, it is probably not going to work very well with ANYCAST, you certainly have no reason to expect it to work at all over ANYCAST. Which in my view makes it a bad discovery scheme.
Since I am always going to do discovery before my application, I see absolutely no reason I would use TCP over multicast. Even if I decide DNS discovery isn't discovering enough, I am going to be making an alternative/supplemental discovery, to a geo-location or network-location sensitive service then doing my application.
TCP over ANYCAST is crossing the streams. Not surprised it doesn't work and see no reason to change that.
On Tue, Aug 3, 2021 at 5:45 AM Vasilenko Eduard <vasilenko.eduard@xxxxxxxxxx> wrote:
