Song: I'll be hosting the meeting
- I'm Song from China Mobile, working with Sudeep (Intel), Dan
(AT&T), Dom (W3C) to run these meetings
... would like to introduce 2 new attendees to this call
... Professor Qiao from BUPT (Beijing Univ Post &
Telecommunications) - top engineering school in
Telecommunication in China
... led the research and develop of the 3G, 4G and 5G in
China
... Profession Qiao heads the network transmission institute
and AC Rep of BUPT
... We also have Qingqian - he used to work for Baidu and now
for a P2P CDN start up and he will give us a guest presentation
today
Sudeep: packed agenda today -
we'll summarize our meetings at TPAC
... which led to the creation of 2 work streams - Dan will talk
about Edge Computing, and I'll talk about Network
prediction
... Song will describe the video cloud service
... and Qingqian will describe his work on a P2P CDN
Song: at TPAC, we had several
relevant evetns
... a full day meeting on Sep 17, demos run on Sep 18 and a
dedicated breakout session on edge computing
... the IG meeting discussed 3 main topics: scope/charter/tasks
review
... guiding principles around the notion of network hints
... discussed use cases and requirements based on input from
participants
... there were over 35 participants at the meeting - the links
to the various presentations were posted on the list
... Key take aways from the meeting: we discussed use cases and
requirements, looked at new solutions like link performance
prediction,
... discussed extending developer tools to support better
network
... We then had a breakout session during the unconference on
edge computing
... discussing several relevant use cases
... There was also a demo of the link performance prediction -
the video will be available soon
DanD: after the TPAC meeting and
the TPAC breakotu session, we concluded that Edge computing is
a particularly interesting topic for several participants
... we want to do more of a deep dive on the topic
... we hear a lot about edge computing in different fora and
industry corners
... edge computing means a lot of different things to different
people
... there are questions about where the edge live, what
constraints apply to it
... the standardization work so far has happened mostly at the
infrastructure level, not at the application level
... which need further exploration
... so we decided to approach this under two different
angles:
... the first is to look at existing edge use cases with high
bandwith / low latency requirements, e.g. AI/ML, Games,
AR/VR
... and look at them from the Web perspective - how they would
apply in the Web context
... we need to address a number of questions - how to make
offloading transparent?
... is offloading visible to the client app or is it hidden by
the browser sandbox?
... e.g. if you are to push some private keys for
authentication - it's one thing if they stay on device, vs
on-prem edge, vs a place where they could be stolen
... power would be another factor
... these are the kind of questions we want to look at in our
first angle on this
... the second aspect was brought up by a number of folks and
is to look at offloading some of the APIs to the edge,
following the model of the split user agent concept
... a particular example would be to push a web worker, a
service worker to the edge
... it could be used in the context of the Immersive Web, Web
of Things
... there are challenages around statefulness, discovery,
sharing of edge instances
... Concretely, we want to engage with key stakeholders to hear
from their ideas
... we will have calls dedicated to this topic - the idea is to
consolidate the ideas from both W3C and non W3C experts towards
a proposed design
Qiao: we could present our
experience on the topic of splitting DNN between mobile browser
and edge
... we've used this in the context of a Web AR
applications
... I can prepare slides on this for a future meetnig
cpn: wanted to briefly mention
work that has happened in the Media & Entertainment
IG
... we had a task force that looked at Cloud browser, who
developed architecture and requirements that might be useful to
this new work
<cpn> https://www.w3.org/2011/webtv/wiki/Main_Page/Cloud_Browser_TF
Sudeep: should this be presented at an upcoming meeting?
cpn: the initial proposers are no longer so active in W3C, so not sure how likely that would be, but will investigate
Sudeep: the second workstream
came out of a very nice presentation from Intel during our TPAC
meeting
... the presentation was on Link Performance Prediction,
discussing how to predict network quality and adjust
... this was well received by IG participants
... we decide to start a workstream on this - there was also
interest from the Intel team to look if and how this could be
brought to browsers
... our goal is to explore how web apps can take advantage of
network metrics and predictions
... we have identified 4 or 5 phases to approach this:
... first, analyse use cases and requirements - we have already
a good starting point
... we learned from TPAC that there is existing work in this
space (Net info API, Media & Entertainment existing work,
...)
... we need to review how much of this matches the need for
this work
... in a second phase, we need to understand what needs to be
done in the browser, in libraries, elsewhere
... next 3 phases depend on the result of the 1st two
phases
... we would want lots of discussions and reviews from privacy
IG, see intersections with WebRTC, Media & Entertainment to
ensure we're in sync
... there will likely be need to coordinate with other
standards groups
... As an IG, we're not doing specs - we could publish a white
paper summarizing our findings
... we could also push for prototypes, maybe interop around
them
Song: I want to talk about the
perspective from operators and developers for cloud service
videos
... Use cases include low-cost and high-efficient live video
broadcasting
... in the Chinese market, there are two obstables for a
in-house video system:
... the cost of of infrastructure resources
... even if you have the structure, it has to remain idle for
most of the time
... One approach that has been taken is to use P2P CDN for
this
... there are 4 top P2P CDN providers in China: Baidu, Alibaba,
Tencent, iQIYI
... from a Web development perspective
... the fundamental needs for a Web-based video service -
capture and codec API and decode / playback
... and a service API to interact with the video cloud, incl
features for streaming, control, playback, analysis and
security
... questions worth considering: how can web developers use
video cloud more easily? would be useful both for cloud
providers and devleopers
... in terms of security monitoring - how can video cloud be
made efficient from a power and cost perspective?
... how do we bring the Web of Things perspective in
this?
... when looking at edge resources deployed by operators in
base stations, how can these be used by developers?
Dom: how would you think we should approach these questions?
Song: as operators, we would love
to see standards from W3C in this space and we would adopt them
in our deployment of 5G
... China Mobile runs an HTML5-based broadcast service across
the country
... on 1st October (a national holiday), we got 1.2 billion
mobile users access this service
... that service also ran multiple views of the broadcast
... this was only possible thanks to Edge computing deployed on
our base stations across the country
... but anyone who would want to do the same would face the
challenges I described
Qingqian: I want to give an
overview of the PCDN (peer-to-peer CDN) technology
... how Tingyu built our solution, and want to explore the
intersection with the Web
... So what is a PCDN?
... a classical CDN system use HTTP and DNS to distribute the
load across server nodes
... a PCDN is not based on DNS - the client contacts a tracker
system (a super node, or a centralized system) which sends all
the edge nodes to the client
... the client can then connect to all the edge nodes
... CDN costs a lot compared to a PCDN which can use idle
upstream bandwidth
... The architecture of the Tingyu PCDN relies on an SDK for
native apps - we're building one of the Web as I will discuss
soon
... the client SDK has two features: it can use the P2P
protocol to link to the edge nodes and download the
content
... THe PCDN network has 3 type of edge nodes: the normal node
like an IoT home device; a core node near the user; a super
node as a central area node (closer to a cloud node)
... we have a decentralized filesystem that can syncrhonize
files across these nodes
... the client SDK can also use HTTP to download the content
from a classical CDN system
... which is itself synchronized with the PCDN decentralized
filesystem
... if the user watns to download the content, the two
protocols can be used at the same time
... the developer uses a server-side API to provide the content
to the network
... I mentioned we've implemented this for native apps - we
want to implement our SDK for Web browsers as well
... but there are open questions when we implement it:
... first, how can the browser connect to the PCDN? we use a
P2P protocol just like bittorrent to connect to the edge node?
can webrtc be used for this? another API?
... once connected, what protocol can we use between the
browser and edge nodes?
... third problem: can we have an efficient way to find edge
nodes by a centralized server or by another way?
Dom: WebRTC sounds like a good canditate - have you looked at Web Transport as a possible new approach?
Qiangqian: we'll try WebRTC - but we lack practice, so it will take time
Sudeep: this whole idea of P2P CDN - it's driven by scalability, not looking at latency, right?
Qiangqian: the goals are similar to a CDN, but with lower costs
Song: low-latency would be another advantage of the proposal - lower cost is probably the primary motivation
sudeep: so it's a bit of a mix of
both
... re discovery of edge nodes, why not reuse DNS for it?
Qiangqian: edge nodes are used to
deliver the content via a P2P protocol (e.g. bittorrent)
... we use DNS for the centralized server
... it could run on HTTP and return the list of edge nodes -
but is there another way?
... we may need a standard way to expose edge nodes via a
standardized structure for edge node list
Song: the slides and the minutes will be sent on the mailing list
Sudeep: we should extract some of these questions to github to track them