Re: More context on ATSSS use case

[David Schinazi <[e-mail]>]

Hi Florin, inline [DS2].

On Sun, Oct 25, 2020 at 4:20 PM Florin Baboescu <
[e-mail]> wrote:

> Hi David. Inline [FFF2].
>
> -Florin
>
>
>
>
>
> *From:* David Schinazi [mailto:[e-mail]]
> *Sent:* Sunday, October 25, 2020 2:03 PM
> *To:* Florin Baboescu <[e-mail]>
> *Cc:* QUIC <[e-mail]>
> *Subject:* Re: More context on ATSSS use case
>
>
>
> Hi Florin, comments inline after [DS].
>
>
>
> David
>
>
>
> On Sun, Oct 25, 2020 at 1:01 PM Florin Baboescu <
> [e-mail]> wrote:
>
> Hi David,
>
> Please see inline [FFF]. Thanks.
>
> -Florin
>
>
>
>
>
> *From:* David Schinazi [mailto:[e-mail]]
> *Sent:* Sunday, October 25, 2020 12:18 PM
> *To:* Florin Baboescu <[e-mail]>
> *Cc:* QUIC <[e-mail]>
> *Subject:* Re: More context on ATSSS use case
>
>
>
> Hi Florin,
>
>
>
> Thank you for your response, please see comments inline.
>
>
>
> David
>
>
>
> On Fri, Oct 23, 2020 at 7:54 PM Florin Baboescu <
> [e-mail]> wrote:
>
> Hi David,
>
>
>
> I see that you are repeating a statement with which I definitely can not
> agree “I'm noticing a pattern where no one is able to explain how this
> will improve the end-user experience though, so I'm going to assume that
> this is beneficial for carriers and not end-users.” So I’ll try to give
> it a try. First I thought it was not necessary as there were already some
> great presentations by Christoph, Olivier and the guys from Alibaba which
> should have provided you with a very good reasoning on benefits for the end
> user. I am not going to go again through what they presented.
>
>
>
> The Apple and Alibaba presentations were great - but they were about
> end-to-end multipath, not about ATSSS. My statement about user experience
> is very specific to ATSSS.
>
>
>
>                We also had one slide in our presentation, which may have
> been overlooked, detailing at least three elements through which a
> multipath access solution may improve the overall Quality of Experience for
> the end user:
>
> 1)      Increased capacity, 2) increased coverage and 3) increased
> reliability.
>
>
>
> Let’s assume for simplicity an user which would be charged for the amount
> of data he/she would use over a cellular network using licensed spectrum
> while for all the data exchanged over the WiFi. While the user is under a
> good WiFi coverage all his traffic is going to be routed over the WiFi, no
> data traffic is going over the cellular. However, when the user is in an
> area of limited coverage or the level of interference reaches a certain
> threshold the quality of the communication over the WiFi access degrades.
> As a result the achievable throughput over the WiFi may get below a certain
> threshold. At this moment the WiFi access may not be able to sustain the
> throughput the user may require. The user may either switch over to the
> cellular (paying a higher penalty) or use both accesses, WiFi and cellular.
> When both accesses are used,  all the traffic below a maximum threshold
> will go over the WiFi access, while all the leftover traffic will go over
> the cellular access.
>
>
>
> Thank you - this is exactly the kind of user benefit I was asking for. In
> other words: the user can save on monetary costs when watching a high
> quality video in this very precise scenario of functional but slow Wi-Fi
> while in range of high speed but expensive cellular. We can debate how
> common this scenario is and how impactful the solution would be, but I
> agree that this is an actual end-user benefit.
>
>
>
> In total for this example there are the following cases:
>
> -          User entirely under the WiFi coverage
>
> -          User entirely under the cellular coverage (no WiFi coverage)
>
> -          User under both WiFi and cellular coverage
>
>
>
> This solution essentially increases the coverage area for the user
> complementing the use WiFi with cellular in zones of poor coverage or no
> coverage.  Without it the user would have been left without data access in
> areas of no WiFi coverage, or with a high rate of error and limited
> throughput access in the areas of poor WiFi coverage or high interference.
> The solution increases the reliability, allowing the user to backup the
> primary access(in this case WiFi) with a secondary access (cellular).
>
>
>
> I don't understand how this increases coverage, can you elaborate?
> Coverage is determined by the cellular antennas and Wi-Fi access points
> available to me. I can use all of them with QUIC connection migration
> end-to-end today, or even plain-old TCP that creates a new connection on IP
> change - ATSSS doesn't change the network coverage I have.
>
>
>
> [FFF] In the example I provided the primary access was over the WiFi
> access. The coverage is increased by augmenting the coverage over the WiFi
> with coverage over the cellular. While I do agree with you that QUIC
> connection migration may be used in certain scenarios when you completely
> switch from one access to another there are still plenty of space left in
> which both accesses may be useful to being used. This scenario is not
> covered by the connection migration. One example in which something like
> this may happen is in areas of high interference or reduce signal strength
> on the primary access, while the secondary access may still be throughput
> limited.
>
>
>
> [DS] Your point here is that there might be value in using two paths. It
> doesn't contradict my point that this does not increase coverage. My
> definition of coverage is "the geographic area that is covered by networks".
>
>
>
> [FFF2] In my opinion the notion of coverage is user dependent and it does
> have both a geographical and a temporal dimension. A user may find himself
> out of the coverage for a service even without movement for various
> reasons. From this point of view I have to admit that I do not quite
> understand your definition of coverage.
>
> The point I was trying to make that the area of coverage for a service for
> a user may be increased by complementing the primary access of the user (in
> my previous example WiFi access) with a secondary access. While I may have
> an excellent coverage in my house using WiFi and I can access any video
> services , as soon as I step outside of my house the quality of my WiFi
> signal degrades rapidly with the distance from the WiFi AP. The application
> becomes rate limited and as a result one can safely say that there is no
> coverage for me for the video service I described at my desired rate.
> However, if I would be able to complement my weak WiFi access outside the
> house with a cellular access, then I may be in the position to be able to
> run my application at the desired rate. In this context I can say that the
> service coverage can be provided outside of the house. Similar argument can
> be brought in the context of the temporal dimension for the coverage.
>

[DS2] Thanks for clarifying. If I understand you correctly, your definition
of coverage is "the geographic area that is covered by networks that
provide throughput higher than X". With that definition, I agree with you
that this can improve coverage in the areas where there is coverage of
multiple networks where none of them individually meet the throughput
requirement but the sum does.


> Similarly, I agree that in some circumstances this can increase overall
> throughput, but I don't understand how it improves reliability, if anything
> I think it'll reduce reliability by adding a new single point of failure in
> the network - the ATSSS proxy.
>
>
>
> [FFF] I’m not sure what do you mean by arguing that we insert a single
> point of failure in the network – the ATSSS proxy. Trying to clarify this
> point let me say, in a very simplistic way that the end to end path between
> the terminal(UE) and a server(S) is made up of the following:
>
> UE <-> access (cellular/WiFi) <-> UPF <-> S
>
> The 3GPP ATSSS solution is based on the notion of a Multi Access PDN. In
> this solution the ATSSS proxy is an on path functionality situated at the
> UPF (aka user plane gateway function) with the goal that together with the
> ATSSS functionality on the UE it accommodates one or more accesses being
> used between the UE and UPF. So from this perspective I do not see how you
> can state that we introduces a single point of failure in the network. On
> the contrary the increased reliability is provided by augmenting one access
> (in my example over WiFi) with a cellular access.
>
>
>
> Let me try to provide you with another example. Let’s assume a client IMS
> based application. The user uses WiFi as a primary access and would like to
> start a high throughput session with a server. On the WiFi side the WLAN
> modem may provide the application with “minimum throughput per access
> class” that can be achieved at a specific moment. Based on this information
> , the application uses IMS to negotiate the end to end pipe to the server
> with a “guaranteed” throughput. While the quality of the access over WiFi
> remains unchanged and the WiFi load is unchanged the “guaranteed” rate can
> be sustained. However, when access characteristics change the value for
> this metric may not be sustainable. In this case augmenting the WiFi access
> with the cellular use addresses this lack of reliability.
>
>
>
> [DS] Let me clarify why ATSSS introduces a new point of failure. First
> here's the common network topology today without ATSSS. In this example the
> Client is a smartphone, and the user is browsing a website hosted by
> WebServer.
>
>
>
> [Client] ----- { cellular } ------- [WebServer]
>
>     \------------{ Wi-Fi   } ------------/
>
>
>
> The client has (at least) two IP addresses: IP_Client_cell and
> IP_Client_WiFi.
>
> The WebServer has IP_Server.
>
>
>
> Now let's assume the Client is happily browsing using QUICv1 over one
> interface, if that network goes down, QUIC will automatically migrate to
> the other network and the web browsing will not fail. The client can send
> packets to IP_Server on any interface and as long as one interface works
> they will reach the server.
>
>
>
> Now let's introduce ATSSS:
>
>
>
> [Client] ----- { cellular } ------- [ATSSS Proxy] -------- [WebServer]
>
>     \------------{ Wi-Fi   } ------------/
>
>
>
> The ATSSS Proxy has IP_ATSSS.
>
> The packets sent by the client are not sent to IP_Server here, they're
> sent to IP_ATSSS.
>
> If the ATSSS proxy becomes unreachable mid connection, the client's
> connection will break - because it's sending to IP_ATSSS not IP_Server.
>
>
>
> My point is that in this scenario ATSSS reduces reliability by adding a
> point of failure.
>
>
>
> [FFF2] It looks like there is a misunderstanding on the ATSSS solution
> here. As I mentioned before the basic element of the ATSSS solution is the
> MA PDN session.  In a MA PDN session there is a single global IP
> address/prefix associated with the client. All the packets sent by the
> client are sent using the client global IP address as a source and the
> destination IP of the server. The ATSSS function resides both at the client
> and at the associated UPF. All the packets of the PDN session go through
> the same associated UPF (which contains the ATSSS functionality). The ATSSS
> function at the UE is provided with two local link IP Addresses which are
> used to establish the MPQUIC connectivity to the ATSSS function at the UPF.
>
> So in short the ATSSS MPQUIC based solution tunnels IP or Ethernet
> packets/frames  between the ATSSS client application on the UE and the
> ATSSS application on the UPF. The transport between the UE and UPF is
> provided via QUIC encapsulation.
>
>
>
> I’ll use the following notation:
>
> IPg = global routable IP address at the UE
>
> IPs = global routable IP address of the Web server
>
> IP1 = local link IP address
>
> IP2 = local link IP address
>
> IP3 = local link IP address
>
> IP4 = local link IP address
>
> *Single Access PDN*
>
> [Client] IPg----- { cellular } ------- [UPF] ------- IPs[WebServer]
>
>
>
> *Multiple Access PDN with MPQUIC ATSSS*
>
> [Client]  IPg  [ATSSS Client] IP1----- { cellular } ------- IP3 [ATSSS
> Proxy][UPF] --------IPs [WebServer]
>
>                                       IP2 \------------{ Wi-Fi   }
> -------------------/IP4
>

[DS2] Thanks for the diagram. In this example, what IP addresses do the
packets for the end-to-end Client-WebServer connection use in both
directions? Because if they're encapsulated in the ATSSS client and
decapsulated in the UPF (and vice versa in the other direction), then if
the UPF becomes unreachable the connection breaks, right?


> Please note that one may think that there may be another deployment
> scenario. *This scenario is NOT covered by the 3GPP ATSSS solution.* For
> this scenario I do agree with you that it may introduce a potential point
> of failure as the ATSSS node may not be on path. In this scenario the user
> creates a single access PDN for Internet and uses the WiFi as a local
> offload. For simplicity of the conversation I’ll limit the case to the the
> terminal having a global routable address over WiFi (IPg2)and a global
> routable address over cellular(IPg1). The ATSSS proxy may also need to
> provide the client with a globally routable IP address reachable from the
> client via the WiFi access. I not this address IPg3.
>
> [Client]  IPg1  [ATSSS Client] IP1----- { cellular } -------[UPF] -------
> IP3 [ATSSS Proxy][UPF] --------IPs [WebServer]
>
>                                       IPg2 \------------{ Wi-Fi   }
> ---------------------------------------/IPg3
>
>
>
> On a side note I would also try to answer a different question. Is the
> bandwidth aggregation solution always useful? Based on various companies
> contributions in 3GPP it was noticed that there is no benefit for the user
> to do bandwidth aggregation when the throughput ratio between the two
> accesses exceeds somewhere between (3-5):1.
>
>
>
> That's a useful datapoint, thanks! So it is restricting the use benefit
> scenario above to only when Wi-Fi and cellular have comparable throughput.
>
>  [FFF] Correct. When we discuss order of magnitude difference using
> bandwidth aggregation does not make any sense. IT is like using a straw
> when your toilet is overrun.
>
> Another interesting use case addresses one of the limitations of WiFi (
> before WiFi6, which uses an OFDMA based access). As many of you know, in
> WiFi an user can transmit only after it detects that there is no one else
> transmitting at the same time. Because of this when the number of users
> served by the same access point increases the quality of the access
> decreases, as all the users compete for the same access. In this case the
> end user may use the WiFi access for all the downlink traffic while for the
> uplink traffic may use the cellular access. This use case improves for the
> end user both the capacity for both downlink and uplink as well as the
> reliability.
>
>
>
> I wouldn't call carrier sense a limitation of Wi-Fi, it's a very
> reasonable design choice that's worked incredibly well for us so far.
> Sending the uplink and downlink packets on different paths sounds like an
> interesting research topic, has this been tried in practice?
>
> [FFF] Yes.
>
>
>
> [DS] Great! I'd love to learn more. Do you know if the data is publicly
> accessible?
>
>
>
> [FFF2] We can take this one offline as I do not think that it belongs in
> this thread.
>
> These are just few examples which try to show the benefits of bringing a
> multipath solution in the toolbox for both end user as well as network
> elements/functions. I hope I brought some more clarity to you.
>
> Regards,
>
> -Florin
>
>
>
>
>
>
>
> *From:* QUIC [mailto:[e-mail]] *On Behalf Of *David Schinazi
> *Sent:* Friday, October 23, 2020 6:12 PM
> *To:* Mirja Kuehlewind <[e-mail]>
> *Cc:* [e-mail]
> *Subject:* Re: More context on ATSSS use case
>
>
>
> Hi Mirja,
>
>
>
> I understand how in some scenarios this could increase throughput.
>
> However, can you clarify how this could improve latency?
>
>
>
> I'm noticing a pattern where no one is able to explain how this will
>
> improve the end-user experience though, so I'm going to assume
>
> that this is beneficial for carriers and not end-users. Unfortunately
>
> I don't have the time to go to 3GPP and do this research myself.
>
>
>
> David
>
>
>
> On Fri, Oct 23, 2020 at 6:07 PM Mirja Kuehlewind <
> [e-mail]> wrote:
>
> Hi David,
>
>
>
> this depends on the actual use case. Using multipath in a masque-like
> proxy setup covers multiple scenarios; in the hybrid access scenario it’s
> throughput, in other cases it can be latency, or a cheaper data
> subscription. That’s what I tried to explain below.
>
>
>
> However, the whole point of ATSSS, as well as other use cases, is to
> provide the (mobile) operator’s costumer/the user better performance that
> what you have right now when using only a single path by actually making
> use of currently unused resources. We can argue what’s the best way to
> achieve that but you probably need to go to 3GPP and have that this
> discussion there. I was mainly trying to explain what ATSSS is, what the
> motivation is, and what the requirements are.
>
>
>
> Mirja
>
>
>
>
>
>
>
> *From: *David Schinazi <[e-mail]>
> *Date: *Friday, 23. October 2020 at 23:08
> *To: *Mirja Kuehlewind <[e-mail]>
> *Cc: *"[e-mail]" <[e-mail]>
> *Subject: *Re: More context on ATSSS use case
>
>
>
> Hi Mirja,
>
>
>
> Can you clarify what you mean by "optimize resource usage and
>
> therefore also the performance for the user"?
>
> 1) What does it mean in networking terms (latency, throughput, etc.)?
>
> 2) What does it mean in end-user terms (video loads faster, etc.)?
>
>
>
> Thanks,
>
> David
>
>
>
> On Fri, Oct 23, 2020 at 12:45 PM Mirja Kuehlewind <mirja.kuehlewind=
> [e-mail]> wrote:
>
> Hi all,
>
> based on the discussion yesterday I would like to provide some more
> context for the ATSSS use case and some notes that probably also applies to
> other proxy based-use cases.
>
> First of all, I would like to clearly note that it's the client (UE) that
> has to request ATSSS support (a Multi-Access (MA)-PDU session) when
> connecting to the mobile network and it's also the client that starts the
> QUIC connection to the proxy (hosted in the UPF). Further for each
> connection that the client starts to some target content server, it can
> again decide to use the ATSSS setup or not (by otherwise connecting to the
> server over a single PDU mobile-network-only session). That means the
> endpoint can locally decide if it wants to only use the mobile link for
> certain connections instead of any kind of ATSSS service. However, that
> decision will likely not only depend on the application characteristics but
> also on e.g. the data subscription, user preferences, or device status.
>
> And that brings me to another point: The right scheduling for the use of
> multiple paths does not only depend on the application characteristics.
> It's also the network conditions of each link, which to some extend can be
> measured in the transport if traffic is sent on both/all links, as well as
> other factors such as user tariff, remaining data volume, or battery
> status. Yes, this doesn't make the problem easier but we also don't need to
> solve this problem in a general way. For each of the proxy-based use cases
> presented yesterday there is a specific network setup with specific
> characteristics and goals. And often the two links do have quite different
> but known characteristics which does make the decision easier.
>
> For the hybrid access case, you have one DSL and one mobile link and
> multipath is used for bandwidth aggregation. This setup is usually deployed
> when the physical line that is serving the DSL doesn't provide sufficient
> bandwidth and in certain areas upgrading those links would be very costly.
> In this case the scheduling is clear: you always fill up the DSL first and
> only use the mobile link when the DSL capacity is exhausted; this can
> happen for e.g. high quality video streaming. In that case the mobile link
> usually has a higher latency and you might need to wait a few more seconds
> before your video starts but I guess that's better than watching the video
> in low quality.
>
> For ATSSS you always have one mobile 3GPP link and one non-3GPP link,
> usually wifi. And as I said in the chat yesterday, for ATSSS this will
> probably get first deployed with managed wifi networks, such that are often
> available today already by mobile operators in certain countries. ATSSS
> also provides a small number of so called "steering modes" which impacts
> the scheduling used, as presented by Spencer yesterday. These modes are
> provided by the network to the client (on the UE) as well as the proxy
> (hosted in the UPF) and these both tunnel endpoints decide independently
> which scheduling to use.
>
> There are different scenarios for these different steering modes, however,
> it's rather a small set of options. When selecting these modes the network
> is able to take additional factors into account such as subscriber data,
> operator configuration, or also application server provided info, e.g. for
> cases where there is actually an SLA between the content provider and
> network operator in place.
>
> By default the scheduling could always prefer one link and only switch
> over when the performance is not sufficient anymore, e.g. the selected
> network gets loaded. While you can measure the network characteristics, and
> ATSSS will also rely on measured characteristics when deciding which path
> to use, the operator of the mobile and wifi networks might actually have
> some additional knowledge about the current network load (number of
> connected user, total traffic volume). Further both the UE as well as the
> UPF in the mobile network might actually have a better view about what's
> happening on the local link than the far end where the content server sits,
> e.g. knowing that a user is moving out of coverage. As such the network
> could for example provide a priority for one path when signaling the
> steering mode and may also indicate certain threshold values that could be
> used to make a switching decision. However, for most flows it might be even
> simpler than that and probably some kind of default mode will be used, e.g.
> based on lowest delay assuming that delay increases when one link gets
> congested.
>
> Another scenario is that a user might choose a cheaper tariff where as
> much as possible of the downlink traffic is off-loaded to wifi. This needs
> to be implemented based on the scheduling in the UPF sitting in the mobile
> network. Further, as the steering modes are provided on a per flow level,
> another example scenarios is that bandwidth aggregation is requested for
> certain traffic flow based on an existing SLA.
>
> Please note that in any of these setups there are multiple e2e connection
> that use the same QUIC tunnel and as just noted each flow can have a
> different steering mode assigned. This is why simultaneous use of both
> paths is especially important for proxy-based use cases.
>
> All these scenarios benefit from knowledge about the local network
> conditions to optimize resource usage and therefore also the performance
> for the user.
>
> Hope that helps,
> Mirja
>
>