Minutes IETF100: nmrg
||Minutes IETF100: nmrg
- Agenda https://datatracker.ietf.org/doc/agenda-98-nmrg/
- Materials https://datatracker.ietf.org/meeting/100/session/nmrg
- Recordings http://ietf100.conf.meetecho.com/index.php/Recordings#NMRG
and on Youtube https://www.youtube.com/user/ietf/videos
* Session 1
Monday, November 13th, Afternoon session III, 17:40 - 18:40 (1 hour)
Scribe(s) and notetaker(s): Thank you Jéferson and Giovane!
1.a) Intent Based Network Management (IBNM)
. Distinguishing Intent, Policy, and Service Models (10 min.),
presenter: Alexander Clemm
MIC Toerless: no definition. scope it and be smaller. change
wrt. existing document (RFC7575). Wiat for the term to
stabilize. MIC Sabine: build some bridges between what has been
proposed some years ago in NFVRG (connectivity), focusing on
how to design an intent grammar nad nby diong that specifying
the scope of intent. Alex: yes. link to prior work. intent
grammar +/- out of scope of this draft. MIC Toerless: not a try
to define anything called intent. Alex: not in the corpus of
the document but references to extensive literature
. Concepts of Network Intent (10 min.), presenter: Mouli Chandramouli
MIC Dean: congested link. well-known problem. what other
problem are in the scope? Mouli: intent is ambiguous. need to
resolve the ambiguity somehow. open problem. MIC Dean: was
asking for other use cases MIC Diego: what the goal of these
documents MIC Jabber/Mouli: one area of research topic that we
want to pursue is network intent. only initial outline
presented now. use cases was mentioned; actual customer use
cases would be useful
. Discussion and Next steps (15 min.)
MIC Dean: network management work flows from operators, eg.
peering, how they translate services into logical/physical
topologies replace a chassis router with physical devices in
different topologies (CLOS), overlays, dimensioning MIC
Expedia: define clearly of the intents. NLP, AI, ... can be
used but don't help if not have a clear problem definition.
real case. going to the stage of writing in common language can
make things worst. AI can help when intent is known (load
balance, anomaly) MIC Toerless: fix terminology. get rid of the
confusion on the semantics of intent. MIC Mouli: digital
assistant (Alexa, Echo, Siri). Two persons ask same question
differently. Look from a networking perspective. MIC Diego:
other term "slice". ok with the work on intent. becoming the
interpretation of IRTF/IETF. important to build/define schema,
information model. how intent, formal description MIC Dean:
cart in front of the horse. first terminology. then use cases
(by operators). business/technical SLA. get that from the
network expressed in simple manner.
Summary of main discussion points:
There are multiple definitions and domains using the
term intent. There are different approaches and means,
different levels/layers, different "producer/consumer".
What is the common denominator: intent as an abstraction.
abstracting the complexity of the underlying layer / network.
So what should the RG do?
Lisandro summarizing the work plan as:
-How to call stuff: --> document(s) on
terminology/taxonomy -How to express/model intent:
information model, grammar, languages -How to realize
Intent Based Network Management System: "reference"
functional architecture, functional blocks/components,
mechanisms/techniques, theory of operations/lifecycle
Additonal work items could also investigate: functions and
techniques ; use cases ; what are the challenges and
corresponing, research items.
1.b) NMRG in 2017, 2018 and beyond (15 min.)
. 2017 RG meetings and progress of work
+ 4 meetings:
- IETF98, Chicago, March. Topic: Autonomic Networks 2.0
- IM2016, Lisbon, May. Topic: New research items
- IETF99, Prague, July. Topic: General meeting, and
Workshop on Measurement-Based Network Management -
IETF100, Singapore, November. Topic: Intent Based
Network Management, and Use of artificial intelligence
(AI) for network management
+ RG document
Autonomic Networking Use Case for Distributed Detection
of SLA Violations
Just passed IRSG poll. Next step: IESG conflict review.
+ 4 active (individual) documents
+ 1 document not published after IESG conflict review.
. 2018 RG meetings plan
*** Tentative list to be discussed ***
+ IETF 101, London, March
+ NOMS 2018, Taiwan, April
+ IETF 102, Montreal, July
+ IETF 103, TBA, November
+ Interim and topical meetings with other groups/events. Open
. Evolution of NMRG
+ Research agenda organized around “themes”
- IBNM, AI techniques, Autonomics 2.0… Other topics of
interest and volunteers to progress the work? - Derive
research items, work plan and milestones
MIC Giovanne: collaborate with MAPRG
+ Do we need a new/revised charter?
MIC Diego: ok with the way forward. don't think need to change
the charter. MIC Alex: good ideas. some themes. not necessary
* Session 2
Tuesday, November 14th, Morning session I, 9:30 - 12:00 (2.5 hours)
Special session on the Use of Artificial Intelligence (AI) for Network
Scribe(s) and notetaker(s): Thank you Jérôme and Giovane!
2.a) Scope and Objectives (5 min.), presenters: NMRG chairs
2.b) Use cases and research results. (~70 min.)
. A deep-reinforcement learning approach for software-defined
networking routing optimization (10 min.), presenter: Albert Cabellos
pros and cons of DRL
lack of explainability -> develop Explainbale Artificial
Intelligence reward function == network management policy MIC
Jabber/Stenio Fernandes: Have you tested your DRL model against
MIC Alex Clemm: How the reward function relates to policy
definition ? How is this defined ?
Albert: The reward function is a continuous function.
The agent tries to increase teh reward function. The
reward function should be the distance between the
current performance of the current and the performance
you want to achieve
MIC ?: Is it proper to SDN cases? Do you have vertical agents ?
Albert: Only one agent on top of the SDN controller. No
MIC ?: What kind of routing algorithms are you using, is it
Albert: Yes it is, The output cannot put "give me the
CLI configuration". The agent will choose the weights
of the links like OSPF and then routers chooses the
paths with the light weight.
. Use of CVAE for Network QoS Management (10 min.), presenter: Shen Yan
pros and cons of CVAE
MIC Yansen: CVAE could help DRL for the training phase
MIC Albert: You predict what will be the QoS given the current
traffic matrix, like a simulator ?
. Network traffic analysis for encrypted traffic and security
monitoring (10 min.), presenter: Jérôme François
MIC/Jabber Stenio Fernandes: question: in your dataset, how
balanced (or imbalanced) is the target class?
MIC Yansen: how relationship between ...?
MIC/Jabber Xiaojin: slide 9
Jerome: classical features from SOTA. and Full feaures:
SOTA + new features on data, and Selected features
MIC Dean: sample rates, applicable on-line? more for forensic
Jerome: near real-time (HTTPS firewall), once HTTPS
session is finished.
MIC Olivier Marjoux: supervised or un- or both
Jerome: signature, HTTPS is supervised. TDA is
MIC Kireeti: how well you can classify?
. Model-free Resource Management of Cloud-based applications using
Reinforcement Learning (10 min.), presenter: Armen Aghasaryan
MIC Dean: what topologies used to test the model?
Armen: simple web applications, parallel load balancers.
Dean: abnormal situations can be topology-specific. The
environment parameters may be dependent to topologies
Armen: avoid the algorithm to perform on abnormal
situations, use other approaches. You don't want to
learn which policies to apply when there is an
anormality. You want to adapt the normal situation
MIC ?/Ericsson: slow convergence, and traffic changes faster
than algorithm convergence. Can you give us what feeling about
convergence time ? what happens if the traffic change faster ?
Armen: if input traffic has no stability patterns, then
no convergence. other approach: maintain Markovian
properties, then convergence can be achieved. No
guarantee about that. To converge, you should assume
that you'll reach a stationarity. The assumption is
that the system behaves like a Markovian model.
Deviations have to be compliant with this Markovian
property. Otherwise, we cannot claim any convergence.
. COGNET project: double closed-loop architecture, and dataset
generation (10 min.), presenter: Diego Lopez
MIC Jabber Stenio Fernandes: how big are these datasets?
Diego: in general, 3 days of traffic. 1-2 TBo
MIC Albert: Did you decide to learn the model offline ?
Diego: initial training is off-line, and then we
MIC: troubleshooting for VNFs should also be an application ?
Diego: Yes. It has been demonstrated by Nokia "noisy
neighbors", as partners of the CogNet project.
. Q&A and discussion (20 min.)
2.c) Emerging landscape of AI in Networks (~50 min.)
. IEEE ComSoc ETC on Network Intelligence (10 min.), presenter: Laurent
Ciavaglia . ETSI ISG Experiential Network Intelligence (ENI) 10 min.),
presenter: Will Liu
MIC Yansen: frequency of operations (rate of data collection).
MEF Open Issues answer regarding collection periods: align the
data for different cope: it depends of the frequency of the
Will: If operation once per second: the higher
frequency data will be useless.
MIC Yansen: ?
. IETF/IRTF Intelligence Driven Networks (10 min.), presenter: Sheng
MIC ?: There are a lot of research in ML, also in network
management like VNF placement among others. This important
topic for which we need data. MIC ?: need the data right, and
control messages to the network.
Sheng: - Sheng: I would to have more data in a
MIC Lisandro: NMRG is a research group. Identify what can be
standardized or not is an important outcome. The
standardization aspect is not the primary end goal.
. Q&A and discussion (20 min.)
2.d) Discussion, conclusions and way forward (~20 minutes)
. What have we learned? What’s important for the next steps?
+ Research items
+ Standardization path(s)
. How do we structure the work to be done
+ Collective roadmap, work plan, deliverables
Diego: importance of having data. An essential tool is to have data.
Dependent of the data you feed the technology will depend the result.
Question about (availability, feasibility of) open dataset? Laurent: in
other domain, there are reference databases (like in image or speech
recognition). Can imagine the same for to have a reference database for
"networking"? Is it a viable approach? or should we rahter seek the
properties of the data we need what are the specific properties these
data should have? Diego: we have old databases, have new ones and keep
them updated. the goal should be similar than in other domains. The
idea would be we should have to keep those data keep them updated
according to the evolution but it is very challenging (data is an
asset). having a set (even not necessarily complete) is important for
reproducible research. Jeferson: based on our previous experience with
Anima WG, a good push should a use case meeting in order to set a draft
(minimum set of use cases). This will give some material to push in the
direction of IA and IDN. Sheng: use cases are important. some already
collected in IDNET. compare use cases, gap analysis. identify from this
future standardization work. Use cases are actually very important and
we collect several ones in the ML. Need gap analysis to know if
something is missing in this use cases. ?: data is important.
methodologies are also important. no reference practice. Data is
important but methodology as well. Traffic classification:
different ways to classify, what insights are needed?
Lisandro: to sum up. In the past, SNMP datasets shown SNMPv3 was not
used as expected. Need an infrastructure for sharing then willingness
to share. Luyuan: focusing on the data is right. open config. real-time
data streaming. how to handle the massive data. (in-memory processing?)
We need real-time data. In top of that you can do your AI. How to
handle the massive data? We need to scale it to reach acceptable
processing time. Compared to other areas which used fixed, our data is
more distributed. Jeferson: importance of use case with default
structure to compare/document. Albert: data sets important. but also an
asset. public dataset used for testing/benchmarking. not for training.
Luyuan: commercial/production data will not be available. Diego: agree.
also unsupersived methods in running networks seems not applicable in
operational networks Luyuan: In our domain, AI is very distributed. One
methodology wil nor solve all problems. You have to target one problem
at a time. Stenio: It is impossible to have data, organize a
competition similar to Kaggle