Re: [Dyncast] CAN BoF issues #7 #17 #32

["[e-mail]" <[e-mail]>]

Hi, Dirk
    
    Please see inline.

Best Regards
Zongpeng Du



[e-mail] & [e-mail]
 
From: Dirk Trossen
Date: 2022-06-17 21:39
To: Dirk Trossen; Linda Dunbar; [e-mail]; dyncast
CC: rtgwg; David R. Oran; jefftant.ietf
Subject: Re: [Dyncast] CAN BoF issues #7 #17 #32
Let me ask my question here a bit more direct: 
 
Is CAN proposing to solve the problem of finding the best compute resources out of possible many possible ones (see item 1 below) or is CAN aiming to do policy-based routing, where such policies may include compute metrics of some sort (which is effectively what I outline in item 2 below)?

[zongpeng]Yes. But the "many possible ones" should have the same anycast address, for example, serviceID1.
                    And CAN will do Load Balancing according to the anycast address.
                    CAN will reuse the policy-based routing mechanism.   [zongpeng]
 
The examples given in the use case and the compared technologies, such as DNS, GSLB, even client-based selection, seem to point towards item 1 but the mentioning of doing path selection is triggering my inquiry here.
 
Dirk
 
From: Dyncast [mailto:[e-mail]] On Behalf Of Dirk Trossen
Sent: 17 June 2022 09:02
To: Linda Dunbar <[e-mail]>; [e-mail]; dyncast <[e-mail]>
Cc: rtgwg <[e-mail]>; David R. Oran <daveoran@orandomnet>; jefftant.ietf <[e-mail]>
Subject: Re: [Dyncast] CAN BoF issues #7 #17 #32
 
Hi Linda, Peng, all,
 
Let us tease apart what “include the path selection” may mean since the nature of this inclusion may be significant in difference.
 
For this, let us assume a service instance S_1 as one of possibly several ones for service S. S_1 may be reachable over a number of network paths, the selection of some of which would significantly impact any compute-aware selection of S_1 over the other available service instances for S. I can see two modes of ‘including path selection”:
 
1.      S_1 exposes two (or more) IP addresses, where each IP address reflects a path from the client to the exposed address. IP addresses may be exposed across more than one network operator, multi-homing the service instance. Now here, ‘path selection’ is indirectly done by picking one IP address over all others, including the IP addresses of other service instances, and indeed, such indirect path selection may well be done through a metric that measures against (at least one) crucial path-related metric. But ultimately, the CAN provider selects one of possibly many IP address still, right? More importantly, it remains the task of the underlay routing infrastructure (again, which could include more than one network operator) to determine what it deems as the ‘best’ path to each of the IP addresses (including the multi-homed S_1 addresses). 
2.      Let’s stick with one IP address to S_1 now though but there are still at least two possible paths to it, where the selection of one over any of the other possible ones could well impact the compute-aware suitability of S_1 over any of the other service instances. Problem here is that ‘including the path selection’ would mean to impact the routing to the single S_1 IP address in a manner that that routing decision takes the compute-awareness into account. The path selection here is not indirect but direct, together with the IP address (i.e., service instance endpoint) selection. What is required here is that CAN provider and underlay somehow work together in selecting one path over another (to the same IP address), which in turn would mean to impact the overall routing decision for S_1’s IP address, which in turn would mean to impact the underlay routing infrastructure since the resulting (compute-aware) path configuration, in the form of suitable forwarding entries, needs distribution in the underlay infrastructure. 
 
I think we have to be clear which of the two options we see in the CAN scope but also if I may have missed options here. As we can see already from those two options, they  have a significant impact on the architecture we may envision for CAN but also for its solution adoption. From my side, I have seen CAN mainly as an endpoint selection problem, so understood ‘path selection’ as an indirect one in the manner described in item 1. I just want to throw the options out here to solicit feedback from the community on this so that we get a good understanding moving forward.
 
Best,
 
Dirk
 
From: Dyncast [mailto:[e-mail]] On Behalf Of Linda Dunbar
Sent: 15 June 2022 23:07
To: [e-mail]; dyncast <[e-mail]>
Cc: rtgwg <[e-mail]>; David R. Oran <[e-mail]>; jefftant.ietf <[e-mail]>
Subject: Re: [Dyncast] CAN BoF issues #7 #17 #32
 
Peng, 
 
For Issue #32, you said: “CAN does not compute path, it selects endpoints.”
 
If CAN means Computing Aware Networking, it should include the path selection. Maybe CAN is about  Selecting (or computing) the optimal paths based on the combination of network conditions and the end point computing available resources? 
 
My two cents, 
 
Linda
 
From: Dyncast <[e-mail]> On Behalf Of [e-mail]
Sent: Monday, June 13, 2022 10:00 PM
To: dyncast <[e-mail]>
Cc: rtgwg <[e-mail]>; David R. Oran <[e-mail]>; jefftant.ietf <[e-mail]>
Subject: [Dyncast] CAN BoF issues #7 #17 #32
 
Dear All,
 
Here are the responses to issues #7 #17 #32, any comments are welcome!  The issues and responses are also copied to the questioner (https://datatracker.ietf.org/doc/minutes-113-can/), hope for further suggestions and confirmation. Thanks!
 
#7 This seems to assume conventional non-distributed applications just running at the edge. What about modern frameworks like Sapphire? and Ray? 
It would be good to understand the multi-site requirements of such frameworks, which seems to mainly run in single DCs.
 
#17 Whether the interests of the organization deploying the application and the organization providing the network connectivity are aligned. Google doesn't worry about this because they are both.
The question is more what the scope and semantic of information is that will need to cross organizational boundaries. This needs further study, in particular when assuming stakeholder division between service and network provider.
 
 #32 How to effectively compute paths? Shall we put CPUs into account? 
CAN does not compute path, it selects endpoints. Path selection (to a given endpoint) is subject to the routing at the IP underlay. For selecting endpoints, CPU information may be taken into account to achieve the 'compute-awareness' that CAN strives for.  
 
You can also add your comments to any of them(https://github.com/CAN-IETF/CAN-BoF-ietf113/issues). 
 
Regards,
Peng
 


[e-mail]
 
From: Linda Dunbar
Date: 2022-05-11 06:11
To: [e-mail]
Subject: [Dyncast] Categories of the CAN BoF issues
CAN BoF proponents:
 
Many thanks for creating the CAN BoF issues tracking  in the Github: https://github.com/CAN-IETF/CAN-BoF-ietf113/issues/created_by/CAN-IETF?page=1&q=is%3Aopen+is%3Aissue+author%3ACAN-IETF
 
I went through the issues captured in the Github and characterized them into groups. Some issues can be lumped together for the discussion. There are quite a few issues related to the requirements, which need to be clarified.
 
Best Regards, Linda
 
 
Issues associated with Applications vs. Underlay networks:
・        Consider not to load underlay network with application details. #35
・        We have multiple upper layer application. Do we have additional needs for routing(e.g. WG?) or we are using those applications and won't need such new WG? #30
・        It needs application information too, so it can't just make a decision at the network layer #23
・        This is not striked as a routing problem; it's all service discovery that can be done in higher layers. #21
・        3GPP and URSP solve this based on UPF selection. It uses both endpoint + application #20
・        One overlay plane per application. Resources/metric specific to the plane. #19
・        How does the application layer or the transport layer learn the network status to steering traffic? #16
 
Need more clear requirements for CAN (to be addressed by draft-liu-dyncast-ps-usecases):
・        Need to understand if three are requirement to avoid extra messages or 1ms of latency #36
・        Regarding the flow affinity, is it from network perspective or from application/computation perspective? #33
・        How to effectively compute paths? Shall we put CPUs into account? #32
・        What happens when the user moves? If so we also need to move application context. #25
・        It can only move the services around as fast as it can update the routing plane. which comes back to the point about service discovery (waiting for convergence/distribution as opposed to just updating the SD server) #24
・        Whether the interests of the organization deploying the application and the organization providing the network connectivity are aligned. Google doesn't worry about this because they are both. #17
o   The question is more what the scope and semantic of information is that will need to cross organizational boundaries. This needs further study, in particular when assuming stakeholder division between service and network provider.
・        It seems impossible to satisfy that requirement simultaneously with the latency requirement #15
・        It wasn't clear that how hard of a requirement session persistence is. #13
o   A session usually creates ephemeral state. If execution changes from one (e.g., virtualized) service instance to another, state/context needs transfer to another. Such required transfer of state/context makes it desirable to have session persistence (or instance affinity) as the default, removing the need for explicit context transfer, while also supporting an explicit state/context transfer (e.g., when metrics change significantly).
・        Should it select UPF based on the application? Steering is done per user? or per application? #9
・        This seems to assume conventional non-distributed applications just running at the edge. what about modern frameworks like Sapphire? and Ray? #7
o   It would be good to understand the multi-site requirements of such framework, which I have understood to mainly run in single DCs.
・        Relation to 3GPP UPF #6
・        Relation to ALTO #5
・        Do the mobility issues and associated protocols are also in scope? There are scenarios where routing alone would not be sufficient. #4
・        What is the position in the edge location regarding to UPF? #3
・        Is there some sort of authorization model so that an edge can indicate whether or not it will provide compute services? #2
・        What is CNC and the relationship with CAN #1
 
Measurement of the Computing Resources (to be addressed by draft-du-computing-resource-representation):
・        It is hard to use existing work to measure the computation, but we can optimize the latency through the performance monitoring. We have performance/measurement matrix over there. #34
・        Clarifications on the computing resource, its requirements and characteristics would be helpful. #27
・        Each application may have a different definition of "resources" these then have to be boiled down into a single topology Network Aware Computing (NAC! :) does scale #14
・        Is computing resource measurable? #10
o   It is, and how to use the measurement would be solution related. See IFIP Networking 2022 paper on how to simply expose “computing capability” and achieve better steering with such simple measure.
・        Why compute resource is different with other resources? #8
・         
Load Balance based solutions:
・        The point is that we need a standardized LB protocol #18
・        The LB as part of the application itself is superior (part of the distributed application itself is to obtain and keep updating the "best" unicast location to use). #22
・        If there is anything missing from current lbs that would prevent their use as-is? other than there is for market reasons no interop standard between different lbs? #12
・        For the load balance, should it learn the network’s status? #11
・         
Dyncast based Solution issues:
・        For Dyncast, when the time is short, is it possible for the router to decide the routing? It is too fast. #31
・        Is dyncast proposed to encapsulate? #29
・        Will CAN dyncast impact each and every router? How to avoid loops? #28
・        What's the assumed scale of a D-router? 10 ^ 6 sessions? 100^ 8? What's the assumed update rate? !Gb? 1Tb? #26