ALTO Virtual Interim 107 Tue, Apr 21 2020 7am - 9am US Central Agenda URL: https://datatracker.ietf.org/doc/agenda-interim-2020-alto-01-alto-01/ YouTube recording URL: https://www.youtube.com/watch?v=yYryHCGvHwM&feature=youtu.be Etherpad URL: https://etherpad.ietf.org:9009/p/notes-ietf-interim-2020-alto-01?useMonospaceFont=true Attendees: 28 (including AD and two chairs) (Raw notes from two note takers are below. Here, we highlight the important discussion in the meeting.) Chairs presented chair slides, agenda bashed, no changes. 4 I-Ds remaining to push to completion: unified-properties, path-vector, performance-metrics, and cdni. cdni and path-vector have a dependency on unified-properties; these drafts will be move forward as a bundle. xdom published as RFC8686. cost-calendar and sse drafts are in RFC Editor Queue. cdni has received reviews from ALTO WG and CDNI WG; need to move the draft ahead. unified-properties appears to be close to being done. Need a WGLC after consultation with authors. performance-metrics appears to need some more list discussion to proceed further. path-vector appears to be close to being done. Need a WGLC after consultation with authors. The remaining discussion in the WG was on individual drafts. There is a lot of interesting and exciting work being proposed, mostly in the "ALTO operational" realm. Details on these talks are in the raw notes below as well as the YouTube recording. Plan of action is to wrap up the four remaining I-Ds by the Madrid IETF. There will be a ALTO meeting in Madrid, virtual or otherwise. ------------------------------------------------------------------------ Raw notes from Qiao Xiang [0] Chair slides: Making progress on quite a few documents. Cost calendar and SSE are in RFC queue. Unified property (UP), CDNI and path vector (PV) receive lots of comments and authors are making quite a few iterations to address the comments and finalize the design. Milestone has been updated. [1] CDNI: Jensen presented that CDNI draft has addressed comments from ALTO WG, CDNI WG and WCLG. Further updates may be needed after the UP draft stabilized. Vijay asked whether CDNI depends on UP. Richard and Jan confirmed. Richard clarified that deleting the part related to UP from CDNI would require too much change, and it may be better to wait for UP to move on together. Martin then asked about whether the dependency is normative. Richard clarified that there is a whole section depending on UP. Jan and Vijay then summarized that CDNI and UP would have to move together since it is not just an RFC number dependency. [2] Unified Property: Sabine presented the summary of changes in v11 of the draft, and the major changes in progress in v12, such as clarifications on entity domain, entity identifier and resource-specific entity property,and the ALTO entity domain type registration. The chairs commented that given the impact of UP on CDNI, it seems reasonable to move these two together. [3] Performance Metrics Richard presented the changes made to this draft. The major change is to introduce the cost-context structure to specify context the metric value is acquired Martin commented that he is inclined to use min/max instead of mean values, because the value is often too dynamic. He also asked is it necessary to introduce RTT as a metric other than one-way latency. Richard clarified that RTT can be different from 2 times one-way latency, and hence RTT is introduced as an optimization. The chairs commented on the question of how to handle statistics of the same metric: Defining a few metric-[] pairs seems to be a reasonable design. Many CSV files use this convention. This design decision does not need to be finalized here, and can be taken to the mailing list. [4] Path Vector Kai presented the changes of the PV draft since v9. The design is stable, but the draft also depends on the UP draft. The chairs asked which part of the PV draft depends on UP the most. Kai clarified that PV needs IANA registry in the UP draft. The chairs further suggested that UP, CDNI and PV authors work closely to move these three drafts together, and UP should close all issues soon since both CDNI and PV heavily depend on it. Richard clarified that authors of these three drafts and a few other WG members are having weekly meeting at 9:30-10:30am US ET every Wednesday to work on these drafts. [5] MoWIE Chunshan Xiong presented Tencent's experience on using mobile and wireless network information exposure to improve applications' performance. Martin commented that this is an interesting presentation, but due to the time, it focuses a little bit too much on application actions. Can min/max of existing ALTO parameters already support these applications? In addition, he is also skeptical on the frequency of updates. Chunshan clarified that traditional ALTO focuses on endpoint selection. In contrast, in these new application scenarios, after endpoints are selected, additional information on the path is also needed to improve applications' performance. He also clarified that the additional traffic load on servers to expose and update network information depends on specific applications. Richard clarified that additional parameters are needed to support mobile applications in the draft. For example, the concurrent number of tower users is not provided by current ALTO. And the current ALTO is not scalable to high-frequency exposure/updated (i.e., every 30 seconds). Finding out the appropriate update interval is an important issue. Ingmar also had a question but due to the time, the chairs suggest moving the discussion to the mailing list. [6] Multi-Domain Danny presented the key questions for ALTO to support multi-domain use cases: what information do multi-domain applications need, what are the issues of gathering such information, and how to design a complete framework, and also presented the current progress in WG, including cross-domain server discovery and several personal drafts. Sebastian asked about the definition of multidomain in this draft. Danny responded that in this draft multi-domain refers to traffic that spans across multiple administrative domains. Ingmar commented on that his team is also working on inter-ALTO-server communication, and suggest that to keep the dialogue open offline. Sebastian further commented that this is an interesting work, and that RFC 8686 has already solved a subset of this problem. [7] BGP-LS Jensen presented the draft of using BGP-LS as an approach to collect the network information for ALTO server. Ingmar commented that his team is doing this already, and different IGPs already have many different representations similar to cost maps. However, mapping traffic to topology is hard and opens a whole can of worms. He suggested keeping the dialogue open offline. [8] Automated Mapping Information Hans presented the problem statement and the fundamental challenge of automated mapping information for ALTO. The key challenge is to determine where flows enter the network. Several designs are discussed, including passive measurement, active measurement and integrating information from ALTo requests. Richard commented that we are interested in on-demand measurement like this, and that Lyle Bertz from Sprint is also studying it. Ingmar commented that his team is also doing the same thing. Vijay commented that it is an interesting work to focus on creating maps in ALTO. 5-6 years ago he also did similar work using US Broadband penetration studies done by FCC to create network maps [R1]. It is interesting to see the efficacy of such solutions when deployed. This is an important operational issue of ALTO. [R1] Vijay K. Gurbani, David Goergen, Radu State, Thomas Engel: Making historical connections: Building Application Layer Traffic Optimization (ALTO) network and cost maps from public broadband data. Proceedings of IEEE CNSM 2014: pp. 193-198 [9] Incremental Update using HTTP/2 Richard presented the motivation on an HTTP/2 based design for ALTO incremental update, and an initial design that maximizes the compatibility. The chairs commented that this is an interesting topic and suggest moving the discussion to the mailing list. [10] Function Delivery Network Zhongxing gave a quick presentation on the function delivery network at Chian Mobile. -------------------------------------------------------------------------------- Raw notes from Sabine Randriamasy ------------------------------------------------------------------ ALTO IETF Virtual Meeting 107 Tue, Apr 21, 2020 0700-0900 US Central Time. ------------------------------------------------------------------ ----- AGENDA 0700 - 0710 10" Chair slides and agenda bash Chairs 0711 - 0721 10" ALTO CDNI request [1] J. Zhang 0722 - 0737 15" ALTO Unified Properties [2] S. Randriamasy / J. Zhang 0738 - 0748 10" ALTO Performance Metrics [3] R. Yang / L. Murillo 0749 - 0759 10" ALTO Path Vector [4] K. Gao 0800 - 0810 10" ALTO Multi-domain Use Cases [5] D. Perez 0811 - 0826 15" ALTO Mobile and W/less Information Exposure (MoWIE) [6] C. Xiong 0827 - 0837 10" Deploying ALTO using BGP LSA [7] J. Zhang/L. Murillo 0838 - 0848 10" ALTO Automatic Generation of ALTO Maps [8] H. Seidel / I. Poese 0849 - 0854 05" ALTO and HTTP/2 [9] R. Yang 0855 - 0900 05" ALTO and FTN [9] S. Yang 0900 *** Session Adjourns *** ----- Attendance: 25 people + area director + 2 chairs WG Chairs: Vijay Gurbani, Jan Seedorf Area Director: Martin Duke ----- Minute takers: Sabine, Qiao? ----- ACRONYMS USED IN THE MINUTES CX Chunshan Xiong DL Danny Lachos HS Hans Seidel IP Ingmar Poese JS Jan Seedorf JZ Jensen Zhang MD Martin Duke RY Richard Yang SK Sebastian Kiesel SR Sabine Randriamasy VG Vijay Gurbani PV Path vector draft UP Unified Properties draft XDOM cross-domain server discovery RFC8686 EG e.g. ------------------------------------------- MINUTES ------------------------------------------- ------ Chair Slides ------ CDNI - Jensen Zhang presents ------------------------------------------- VG: is the idea to move CDNI and unified properties (UP) together ? RY: FCI is really dependent on UP there may be independency but this would require too much changes VG if it depends on UP then fine. No need to lobotomize CDNI. Just a process question RY: so we keep dependency VG we just move both together MD: is the issue about text or other? AW: there is a whole section using UP and dependent on UP JS: as a co-author, we wait for UP and publish both together with reference to the correct RFC nb. MD: if the only issue is on having a normative ref then it is OK. If change in (UP?) draft then we need to wait, given the dependency RY: there is chance we need change text wrt UP updates, so we need wait a little. VG: then we want move these together ----- UP - Sabine Randriamasy presents ------------------------------------------- VG: given impact on other drafts. Seems reasonable to have UP progressing with other drafts ----- APM - Richard Yang presents ------------------------------------------- Most changes are in acquisition conditions of metrics. MD : I would recommend using min and max values instead of mean values, because values are often too dynamic. Do we expect RTT be diffirent from to 2 times one way delay as defined elsewhere.? RY: can be up to 30%. VG: RY should push these questions to mailing list VG: defining metric stat pairs is reasonable: eg define metric + min or max and specify a few of them. For machine learning ML metrics, many similar labels are used, EG - moment, min. We can take this decision to the mailing list ----- Path Vector - Kai Gao presents version in progress ------------------------------------------- - no constraints are supported in PV response - reformulating Server response, as the order of parts in the multipart response can be changed VG: this also depends on UP. Beyond that, do authors feel it's baked? Kai: we need to solve IANA issues then we are mostly ready RY: PV, UP, CDNI authors hold weekly meetings. UP dependency slows the work down. If people have expertise, please join on meetings on Wednesday. VG: yes, we need to move. Please people do participate to meetings. Please UP move ahead. ----- JS: for the rest of the presentations please use 80% of your allocated time ----- MD use cases Danny Lachos - not to be heard. Muted? Skip to MOWIE ----- MOWIE for network-aware applications - Chunshan Xiong presents ------------------------------------------- Apps use many user and server measurements to adapt QoE impacting params. Measurements are near real time and not precise. Standards for info exposure exist in 3gpp. MD: thanks. Interesting work? A little too focused over the application and what it does with data. Curious to see if you can use ALTO metrics such as min-max BW. How dynamic does it need to be? CX: MOWIE differs with ALTO. ALTO selects endpoints to save transport resource. In MOWIE, after the path is selected, we need to monitor path to provide good QoE. MD: network info exposure can cause additional traffic. Ingmar's question will be taken to the list VG: please send comments on list RY: comment on Martin's question on whether max-min BW is good enough. The number of tower users also necessary and not provided by ALTO. Implication of real time data needs a flexible way to expose performance metrics. If ALTO can push in every 30 secs, can we 10 msecs? ----- Multi-domain use cases - Danny Lachos presents ------------------------------------------- SK: what is multi-domain? the original ALTO use case was BitTorrent which is multi-domain. Whats new here? IP: beware we work with several ALTO servers. If you want to get info and results, we are open for discussions SK: interesting work. A subset is already adressed in XDOM that aims at parts of your work. I am interested to discuss on what's the gap. ----- BGP LS - Jensen Zhang presents ------------------------------------------- BGP LS is a potential candidate to collect info on inter-domain topology, CIDR distrib, routing info, perf metrics. Cannot talk with all BGP router, so we need filtering requirements to select the relevant ones. We present early design of mechanism to do this IP: Did you think about considering that different implems exist to compute and represent cost maps? We work on this pb since 3-5 years. We can discuss. Topology however gives no clue on flows. JZ: we only consider... IP you want actually to map traffic and not topology. VG: there is many existing work on how to create maps. ALTO maps were initially meant to be computed by providers. We also have worked on this, using public broadband data and I will post the references. ----- Automatic map generation for ALTO - Hans Seidel presents ------------------------------------------- Challenge: tracking flows. Router does not tell much outside the network view. Other problem: most traffic comes from outside the network. One option is passive measurements to detect ingress points, using netflow, sflow etc. but is very heavy. Gbits of data per sec. Other option is active measurement by content provider with Traceroute. RY: there is strong interest for on-demand map, a key person in Lyle Bertz from Sprint. ----- SSE with http2 - Richard Yang presents ------------------------------------------- IESG review on SSE asked about using HTTP2 SSE uses http 1.1 and allows sending only 1 request at a time. An HTTP2-based design has started, studying features for backwards compatibility. ----- Function Delivery Networks (FDN) - no slides - Zhong Xiang presents ------------------------------------------- FDN similar to CDN. It does not deliver content but functions. FDN is deployed in the bChina Mobile network. We integrate the idea of ALTO for more efficient deployment. We use IBM openwhisk and KB to dissemintate dockers. ----- WG Chair wrap-up ------------------------------------------- JS: thanks a lot to everybody to participate. We want to close milestones by Madrid. VG: see you in Madrid, virtual or real.