| Internet-Draft | QoO | February 2023 |
| Olden & Teigen | Expires 7 August 2023 | [Page] |
This document describes a new network quality framework named Quality of Outcome (QoO). The QoO framework is unique among network quality frameworks in satisfying all the requirements layed out in "Requirements for a Network Quality Framework Useful for Applications, Users and Operators".¶
The framework proposes a way of sampling network quality, setting network quality requirements and a formula for calculating the probability for the sampled network to satisfy network requirements.¶
This note is to be removed before publishing as an RFC.¶
The latest revision of this draft can be found at https://domoslabs.github.io/QoOID/draft-olden-ippm-qoo.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-olden-ippm-qoo/.¶
Discussion of this document takes place on the IP Performance Measurement Working Group mailing list (mailto:ippm@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/ippm/. Subscribe at https://www.ietf.org/mailman/listinfo/ippm/.¶
Source for this draft and an issue tracker can be found at https://github.com/domoslabs/QoOID.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 7 August 2023.¶
Copyright (c) 2023 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
"Requirements for a Network Quality Framework Useful for Applications, Users and Operators" [draft-teigen-ippm-app-quality-metric-reqs] describes a set of requirements for a network quality framework. This document explores how the quality attenuation metric and framework [TR-452.1] can be extended to meet the full set of requirements.¶
Quality attenuation is a network quality metric that meets the most of the criterias set out in the requirements; it can capture the probability of a network satisfying application requirements, it is composable, and it can be compared to a variety of application requirements. The part that is yet missing is how to present quality attenuation results to end-users and application developers in an understandable way. We believe a per-application, per application-type, or per-SLA approach is appropriate here. The challenge lies in specifying how to simplify enough without losing too much in terms of precision and accuracy.¶
We believe the probabilistic approach is key as the network stack and applications network quality adaptation can be highly complex. Applications and the underlying networking protocols makes separate optimizations based on their perceived network quality over time and saying something about an outcome with absolute certainty will be practically impossible. We can however make educated guesses on the probability of outcomes.¶
We propose representing network quality as minimum required throughput and set of latency percentiles. Application developers, regulatory bodies and other interested parties can describe network requirements in the same manner. We propose a formula for a distance measure between perfect and useless quality. This distance measure can, with some assumptions, calculate something that can be simplified into statements such as "A Video Conference has a 93% chance of being lag free on this network" all while making it possible to use the framework both for end-to-end test and analysis from within the network.¶
The work proposes a minimum viable framework, and often trades precision for simplicity. The justification for this is to ensure adoption and usability in many different contexts such as active testing from applications and monitoring from network equipment. To counter the loss of precision, we require some parameters that allow for analysis of the precision.¶
The foundation of the framework is Quality Attenuation [TR-452.1]. This work will not go into detail about how to measure Quality Attenuation, but some relevant techniques are:¶
Quality Attenuation represents quality measurements as distributions. Using Latency distributions to measure network quality is nothing new and has been proposed by various researchers/practitioners. The novelty of the Quality Attenuation metric is to view packet Loss as infinite (or too late to be of use e.g. > 3 seconds) latency [TR-452.1].¶
Latency Distributions can be gathered via both passive monitoring and active testing. The active testing can use any type of IP traffic. It is OSI Layer and network technology independent, meaning it can be gathered in an end-user application, within some network equipment, or anywhere in between.¶
A key assumption behind the choice of latency distribution is that different applications and application categories fail at different points of the latency distribution. Some applications, typically downloads, have lenient latency requirements. Video Conferences typically are sensitive to high 90th percentile latency and to the difference between the 90th and the 99th percentile. Online gaming typically have a low tolerance for high 99th percentile latency. All applications require some level of throughput and packet loss rate. A network quality metric that aims to generalize network quality must take the latency distribution, throughput, and packet loss into consideration.¶
Two distributions can be composed using convolution [TR-452.1].¶
To reach the design goal of being useful in the contexts laid out in "Requirements for a Network Quality Framework Useful for Applications, Users and Operators" [draft-teigen-ippm-app-quality-metric-reqs], this work imposes no requirement on the time period or the network loading situation. This choice has pros and cons. Latency under load is extremely important, but average or median latency has a role too. However, a network quality metric that does not take latency under load into account is bound to fail at predicting application outcome.¶
This framework only requires a latency distribution. If the sampling is done while the network is loaded, latency under load will be part of the distribution, which is encouraged, but is not always possible, for example when passively monitoring the latency of real traffic.¶
It takes quite a few samples to have a statistically significant distribution. Modeling a distribution may be a challenging software engineering task, hence we need to sample the latency distribution at certain percentiles. A list of 10 percentiles in a logarithmic-esque fashion has already been suggested in industry [0th, 10th, 25th, 50th, 75th, 90th, 95th, 99th, 99.9th, 100th] and seems adequate. We propose to define a shared set of percentile values to report.¶
The framework is flexible when it comes to the direction of traffic that is being sampled, but does require that it is noted whether the latency distribution is measured one-way or two-way. The framework does not require an explicit bandwidth measurement, but does require a note on the maximal observed throughput in the time period.¶
By not requiring a specific number of samples, this framework allows taking 10 samples and calling it a distribution, which of course is not ideal. On the other hand, making the framework overly complex and difficult to adhere to using real-world equipment and applications is the best way to ensure that this framework goes unused. Constraints will vary for different network equipment and applications.¶
To make sure we can trust measurements from others and analyze their precision, we require:¶
Type of measurement:¶
By requiring the report of these variables, we ensure that the network measurements can be analyzed for precision and confidence.¶
This work builds upon the work already proposed in the Broadband Forum standard called Quality of Experience Delivered (QED/TR-452) [TR-452.1]. In essence, it describes network requirements as a list of percentile and latency requirement tuples. In words, a network requirement may be expressed as: The network requirement for this app quality level/app/app category/SLA is "at 4 Mbps, 90% of packets needs to arrive within 100 ms, 100% of packets needs to arrive within 200ms". This list can be as simple as "100% of packets need to arrive within 200ms" or as long as you would like. For the sake of simplicity, the requirements percentiles must match one or more of the percentiles defined in the measurements, i.e., one can set requirements at the [0th, 10th, 25th, 50th, 75th, 90th, 95th, 99th, 99.9th, 100th] percentiles. The last specified percentile marks the acceptable packet loss. I.e. if the 99th percentile is 100, and the 99.9th or 100th percentile is not, 1% packet loss (100-99) is inferred.¶
Applications do of course have throughput requirements. With classical TCP and typical UDP flows, latency and packet loss would be enough, as they are bound to create some latency or packet loss when ramping up throughput if subsequently they become hindered by insufficient bandwidth. However, we cannot always rely on monitoring latency exclusively, as low bandwidth may give poor application outcomes without necessarily inducing a lot of latency. Therefore, the network requirements should include a minimum bandwidth requirement.¶
Whether the requirements are one-way or two-way must be specified.¶
Until now, network requirements and measurements are what is already standardized in BBF TR-452 (aka QED) framework [TR-452.1]. The novel part of this work is what comes next. A method for going from Network Requirements and Network Measurements to probabilities of outcomes, or Quality of Outcomes if you will.¶
To do that we need to make articulating the network requirements a little bit more complicated. A key design goal was to have a distance measure between perfect and useless, and have a way of quantifying what is 'better'.¶
We extend the requirements to include the quality required for perfection and a quality threshold beyond which the application is considered useless.¶
Network Requirements for Perfection (NRP), for example: At 4 Mbps, 99% of packets need to arrive within 100ms, 99.9% within 200ms (implying that 0.1% packet loss is acceptable) for the outcome to be perfect. Network Requirement points of uselessness (NRPoU): If 99% of the packets have not arrived after 200ms, or 99.9% within 300ms, the outcome will be useless.¶
Where the NRPoU percentiles and NRP are a required pair. I.e., if the 99.9th percentile is part of the point of uselessness then the network requirements must also include the 99.9th percentile.¶
At this point we have everything we need to calculate the quality of the application outcome. The QoO. There are 3 scenarios:¶
1 and 2 require nothing more from the framework. For 3, we will now specify the calculation between to translate these distances to a 0 to 100 measure. We use the percentile pair where the measured latency is the closest to the NRPoU as the application is only as good as its weakest link.¶
Mathematically: QoO = min(ML, NRP, NRPoU) = (1-(ML-NRP)/(NRPoU-NRP)) * 100)¶
Essentially, where on the relative distance between Network Requirement for Perfection (NRP) and Network Requirement Point of Uselessness (NRPoU) the Measured Latency (ML) lands, normalized to a percentage.¶
Where: NRP is network requirements for perfection. With minimum throughput and with percentiles and milliseconds NRPoU is the points of uselessness. With percentiles and milliseconds i is the length of NRP / NRPoU latency per percentile requirements ML is Measured Latency in percentiles and milliseconds¶
NRP: 4 Mbps {99,250},{99.9,350} NRPoU: {99,400},{99.9,401} Measured Latency: .... 99p = 350ms, 99.9p = 352ms Measured Minumum bandwidth: 32 Mbps / 28Mbps¶
Then the QoO is defined: QoO = Min( ((1-(350 - 250)/(400-250))100),((1-(352 - 350)/(401-350))100)) = Min (33.33,96.08) = 33.33¶
In this example, we would say: This application/SLA/application category has a 33% chance of being lag-free on this network¶
A key advantage of having a measurement that stretches between perfect and useless, as opposed to having a binary (Good/Bad) or other low resolution (Superbad/Bad/OK/Great/Supergreat) metrics, is that we have some leeway. The leeway is useful, for instance: a lower than 20% chance of lag free experience is intuitively not good and a greater than 90% chance of lag free experience is intuitively good - meaning we don't have to find perfection for making the QoO metric useful.¶
Nevertheless we have to find some points for uselessness and perfection. There is no strict definition of when the network is so bad that the application is useless. For perfect, we may have a definition for some apps, but for apps like web browsing and gaming, lower latency is simply better. But to assist those who wish to make a requirement, we can say that if the end-user experience does not change when reducing the latency, the network quality is sufficient for the Network Requirements for Perfection (NRP) .¶
Someone who wishes to make a network requirement for an application in the simplest possible way, should do something along these lines.¶
Someone who wishes to find sophisticated network requirements might proceed in this way¶
A QoO score at 94 can be communicated as "Johns iPhone has a 94% chance of lag-free Video Conferencing", however, this does not mean that at any point of time there is a 6% chance of lag. It means there is a 6% chance of experiencing lag during the entire session/time-period, and the network requirements should be adjusted accordingly.¶
The reason for making the QoO metric for a session or time-period is to make it understandable for an end-user, an end-user should not have to relate to the time period the metric is for.¶
Microsofts own video-conferencing service requirements for Teams can be translated into the QoO Framework. For best performance for video meetings they specify 4/4 Mbps, 100 ms latency, <1% packet loss, and <30 ms jitter. This can be translated to an NRP (if we take some liberties with interpreting their jitter score):¶
NRP Microsoft Teams: At minimum 4/4 Mbps. {0p=70ms,99p=100ms}¶
For minimum requirements Microsoft does not specify anything, but at 500ms latency or 1000ms 99p latency, a video conference is very unlikely to work in a remotely satisfactory way.¶
NRPoU {0p=500,99p=1000ms}¶
Of course, it is possible to specify network requirements for Teams with multiple NRP/NRPoU, for different quality levels or one/two way video and so on. Then one can calculate the QoO at each level.¶
We have described a way of simplifying how the network requirements of applications can be compared to quality attenuation measurements. The simplification introduces several artifacts that may or may not be significant. If new information emerge that indicate other tradeoffs are more fit for our purpose, we should switch before this Internet Draft moves further. In this section we discuss some known limitations.¶
These two latency series: 1,200,1,200,1,200,1,200,1,200 and 1,1,1,1,1,200,200,200,200,200 Will have identical distributions, but may have different application performance. Ignoring this information is a tradeoff between simplicity and precision. To capture all information necessary to perfectly capture outcomes we are getting into extreme computational complexity. As an application's performance is bound by how the developers change to varying network performance, meaning nearly all different series of latencies may have different application outcomes.¶
Additionally, we cannot capture latency on every packet that is sent. We can probe and sample, but there will always be unknowns. We are now in the realm of probability. Perfection is impossible, but instead of denying this, we should embrace it, which is why talking about the probability of outcomes is the way forward.¶
One can conjure up scenarios where 50ms latency is actually worse than 51ms latency as developers may have chosen 50ms as the threshold for changing quality, and the threshold may be imperfect. Taking these scenarios into account would add another magnitude of complexity to determining network requirements and finding a distance measure (between requirement and actual measured capability).¶
Choosing this is to reduce complexity, but we do acknowledge that the applications are not that simple. The defence for this trade off is that insufficient bandwidth will cause queues and therefore latency, and it should be possible to see this. Additionally, network requirements can be set up per quality level (resolution, fps etc.) for the application. However, having too many network requirements also increases the complexity for users of the framework, and it is still unclear if this is the optimal tradeoff.¶
To ensure simplicity, packet loss is described as infinite latency and the resolution will be bound to the percentiles we chose to sample. There is a good argument that some applications need higher resolution on packet loss for sufficiently describing application outcomes. If this good evidence is presented for this, packet loss should be measured separately and added to the QoO formula.¶
There is a need for a selection of percentiles, as we in the name of simplicity can't use them all. But how should we select them? The 0th (minimal) and 50th (median) percentile have implicit usage by themselves. [BITAG] discusses that the 90th, 98th and 99th percentiles are key for some apps. In general the wisdom is that the higher percentiles are more useful for interactive applications, but only to a certain point. At this point an application sees it as packet loss and may adapt to it. Should we pick the 95th, 96th percentile, the 96.5th or the 97th? We don't know, and as this is likely not universal across applications and applications classes, we simply have to choose arbitrarily, and to the best of our knowledge.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
TODO Security¶
This document has no IANA actions.¶
TODO acknowledge.¶