I am wondering whether building Bottleneck structure can be seen as modeling technology and require AI and machine learning technologies to be involved? If the answer is yes, it might be worth exploring this in NMRG mailing list and provide input to AI for network management work item in the NMRG.
Qin correct, bottleneck structures provide a scalable model of the network that can/will be further extended using ML. The NMRG has the following objectives in its charter: 1) self-driving/-managing networks, 2) intent-based networking and 3) artificial intelligence in network management. These are all areas where we believe bottleneck structures can have an impact, so I would be interested in exploring how we can help the NMRG charter. (At the bottom of this page you will see a tree<https://www.reservoir.com/gradientgraph/try-it/> illustrating some of the applications of bottleneck structures; as shown, modeling and ML plays an important role. See also [1]). I created a ticket for this task that I am happy to take with the collaboration of others who might also be interested: https://github.com/ietf-wg-alto/wg-materials/issues/13 Secondly, I am very interested how ALTO can be integrated with PCE? How ALTO can be used together with PCE for path computation and setup, How ALTO exposed information can be better consumed by the client, either in network layer or transport layer or application layer? I agree, I think this is a very important use case for ALTO. We actually plan to make this one of the hackathon's use cases. we had already created a ticket around this: https://github.com/ietf-wg-alto/wg-materials/issues/12 . We can discuss this topic together during our weekly calls and also with the hackathon team. Third, it might be also related to modeling technologies, I am wondering how flow information can be correlated and aggregated into network topology information and other ALTO information such as performance metrics? Is there any mechanism or Guidance or principle, best practice on how difference data source can be integrated together based on implementation experience. Maybe Danny and Jordi can help clarify this from your experience, thanks. In our G2 deployments we use a combination of sources such as NetFlow, sFlow, topology and routing information, etc (see for instance [2]). Bottleneck structures are an example of object that correlates (1) flow information, (2) topology and (3) routing to create a single graph model of the network. The model allows you to make efficient calculations of optimized traffic engineering decisions prior to deploying them onto the production network. For instance, a PCE could use the bottleneck structure to compute an optimized path that avoids congestion. Thanks, Jordi [1] "Designing data center networks using bottleneck structures," ACM SIGCOMM 2021 [2] "Computing Bottleneck Structures at Scale for High-Precision Network Performance Analysis,", IEEE/ACM SC 2020 ________________________________ From: Qin Wu <[email protected]> Sent: Sunday, April 10, 2022 13:08 To: Jordi Ros Giralt <[email protected]>; [email protected] <[email protected]> Cc: Danny Lachos <[email protected]>; Y. Richard Yang <[email protected]> Subject: RE: I-Draft discussion - Supporting Bottleneck Structure Graphs in ALTO: Use Cases and Requirements WARNING: This email originated from outside of Qualcomm. Please be wary of any links or attachments, and do not enable macros. Thanks Jordi for bringing this up on the list for further discussion. I am wondering whether building Bottleneck structure can be seen as modeling technology and require AI and machine learning technologies to be involved? If the answer is yes, it might be worth exploring this in NMRG mailing list and provide input to AI for network management work item in the NMRG. Secondly, I am very interested how ALTO can be integrated with PCE? How ALTO can be used together with PCE for path computation and setup, How ALTO exposed information can be better consumed by the client, either in network layer or transport layer or application layer? Third, it might be also related to modeling technologies, I am wondering how flow information can be correlated and aggregated into network topology information and other ALTO information such as performance metrics? Is there any mechanism or Guidance or principle, best practice on how difference data source can be integrated together based on implementation experience. Maybe Danny and Jordi can help clarify this from your experience, thanks. -Qin 发件人: alto [mailto:[email protected]] 代表 Jordi Ros Giralt 发送时间: 2022年4月8日 19:48 收件人: [email protected] 主题: Re: [alto] I-Draft discussion - Supporting Bottleneck Structure Graphs in ALTO: Use Cases and Requirements Copying over some questions that Qin asked in a separate conversation and including my answers: Regarding section 3, I am thinking whether this require interaction between overlay and underlay, one typical example is SDWAN use case described in https://datatracker.ietf.org/doc/html/draft-dukes-spring-sr-for-sdwan which require SDWAN controller and SR controller better interaction. This is a good question and we had a similar conversation with Luis during the IETF Meetings. In order to construct the bottleneck structure of the network, we need flow path information (i.e., the set of links traversed by the flows). In overlay/underlay networks like SDWAN, ground-truth flow-path information needs to be obtained from the underlay. This information can be derived from protocols like NetFlow / IPFIX. Alternatively, it can also be obtained from the traffic engineering database (TED) or from the SDN controller. In section 4.4 “Optimal Joint Congestion Control and Routing<https://giralt.github.io/draft-ietf-alto-gradient-graph/draft-giraltyellamraju-alto-bsg-requirements.html#name-optimal-joint-congestion-co>”, I assume Routing is at the network layer while congestion control is at the transport layer, do we have similar interaction between network layer and transport layer. I assume some emulation platform you should build to test impact of new added flow on the network. Good question too. In this case there is no need for network and transport to interact. Once the bottleneck structure of the network has been computed, we can jointly solve routing and congestion control. For instance, a PCE could use the bottleneck structure to compute a throughput optimal SRv6 path and then program a flow's packet header SIDs accordingly. See Section '5. Application Layer Traffic Optimization using Bottleneck Structures'<https://datatracker.ietf.org/doc/draft-giraltyellamraju-alto-bsg-requirements/> for an example. Thanks, Jordi ________________________________ From: alto <[email protected]<mailto:[email protected]>> on behalf of Jordi Ros Giralt <[email protected]<mailto:[email protected]>> Sent: Thursday, March 24, 2022 21:36 To: [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> Subject: [alto] I-Draft discussion - Supporting Bottleneck Structure Graphs in ALTO: Use Cases and Requirements WARNING: This email originated from outside of Qualcomm. Please be wary of any links or attachments, and do not enable macros. Hello all, During the ALTO session, the chair asked that we bring to the mailing list the discussion about the new draft "Supporting Bottleneck Structure Graphs in ALTO: Use Cases and Requirements" https://datatracker.ietf.org/doc/draft-giraltyellamraju-alto-bsg-requirements/ I will start by following up on the questions raised during the session via the Jabber chat, but feel free to raise any other questions during this conversation: [QW] Can alto provide sufficient information to build bottleneck structure graph? is there data translation needed? in the scope of your draft or not? [RY] There are two interpretations to this question: (1) build bottleneck structure on top of alto and (2) make bottleneck structure a (new) service provided by alto. For (1), we need some modifications; I think the focus here is (2). Jordi? [JRG] Right, I think the focus is (2). To construct the bottleneck structure, we need information about the set of links traversed by the flows and the link capacity. Flow information can be obtained from protocols such as NetFlow or sFlow (this approach is similar to how other deployments discussed in ALTO collect flow info, such as the Flow Director<https://people.csail.mit.edu/gsmaragd/publications/CoNEXT2019/CoNEXT2019.pdf>). Link capacity information can be obtained from protocols such as SNMP, the SDN controller or topology files. In two deployments in the US we use NetFlow, sFlow and topology files. Information could also be pulled from other sources. For instance, today we discussed with Luis that if PCE is used, we could get path information from the traffic engineering database (TED). Perhaps Luis also wants to add to this discussion. Thanks, Jordi
_______________________________________________ alto mailing list [email protected] https://www.ietf.org/mailman/listinfo/alto
