Re: [Rift] comments on draft-przygienda-rift-dragonfly-00

[[e-mail]]

The group is like pod, fabric or something like. For example in the figure, SA1, SA2, LA1 are in one group. SB1, SB2, LB1 are in the other group.
If the congestion is considered, the computation may be different. But we can consider it later. 
Best regards,
Sandy





 




Original
 




From:prz<[e-mail]>


To:张征00007940;


Cc:rift<[e-mail]>;


Date:2023-11-02 22:22:01


Subject:Re: [Rift] comments on draft-przygienda-rift-dragonfly-00






 









just one default, what do you mean by “group” ?



the draft talks nothing about some kind of “congestion”, it’s completely orthogonal topic. those are new mechanisms/metrics
 that would need to be used 




 


dragonfly draft talks purely about correct reachability advertisement/computation on dragonfly/sparse dragonfly topology with 1-hop alternates


 











Juniper Business Use Only




From: [e-mail] <[e-mail]>

Date: Thursday, 2 November 2023 at 07:35

To: Antoni Przygienda <[e-mail]>

Cc: [e-mail] <[e-mail]>

Subject: Re: [Rift] comments on draft-przygienda-rift-dragonfly-00






[External Email. Be cautious of content]


 




Hi Tony, 


Yes. I also think RIFT protocol is good to be used in dragonfly+ topology. 


And the basic logic of protocol is not changed.  


I have some comments on the 00 version: 


1, If the ToF node still advertises the default route to leaf nodes? If it is, how many routes will be advertised to the leaf nodes, if these routes will be different due to the group?


2, If I understand right, in case the topology is used for AI training, the leaf node will select another route when the used route is congested. So in section 4.1, route needs to be changed when
 the destination is unreachable or congested. 


And futhermore, the dragonfly+ dense and sparse modes are different, will we consider the dense mode also? 


What do you think of it?


Best regards,


Sandy





















Original







From: AntoniPrzygienda <[e-mail]>







To: 张征00007940;







Cc: [e-mail] <[e-mail]>;







Date: 2023年10月27日
 21:04







Subject: Re: [Rift] comments on draft-przygienda-rift-dragonfly-00









daft explains how horizontals are used and that the DF+ backbone is basically collection of special inter-fabric horizontal links


 

draft already adds some procedures for LIE processing and additional computation .

 

Good thing is it does NOT change really the flooding scopes/rules since those are the hardest part of the spec to implement correctly as is usual the case for IGPs


 

if you have alternate proposal, well, there is a WG meet in 2 weeks
😉


 





tony






 







 
Juniper Business Use Only


From:
[e-mail] <[e-mail]>

Date: Friday, 27 October 2023 at 11:40

To: Antoni Przygienda <[e-mail]>

Cc: [e-mail] <[e-mail]>

Subject: [Rift] comments on draft-przygienda-rift-dragonfly-00




 



[External Email. Be cautious of content]

 



Hi Tony, 

Thank you for providing this draft! 

IMO it's an interesting topic that using RIFT in dragonfly or dragonfly+ topology. 

Could you please post the draft figures here for reading the draft easily?

And will you consider using east-west connection between the TOFs? 

Though it will change the flooding and computation procedure in RIFT protocol, we think it may be another feasible way. 

Thank you very much!

Best regards,

Sandy