IAB workshop report: Measuring Network Quality for End-Users

The three day workshop was broken into four separate sections, including introductory material, that each played a role in framing the discussions. This was followed by a discussion about conclusions that could be agreed upon by workshop participants ().

The workshop started with a broad focus on the state of user Quality of Service (QoS) and quality of experience (QoE) the Internet today. The goal of the introductory talks was to set the stage for the workshop by describing both the problem space and the current solutions in place and their limitations. The introduction presentations by participants provided views of existing QoS and QoE measurements and their effectiveness. Also discussed was the interaction between multiple users within the network, as well as the interaction between multiple layers of the OSI stack. Some existing measurement work was also presented. Vint Cerf provided a key note describing the history and importance of the topic.

We may be operating in a networking space with dramatically different parameters compared to 30 years ago. This differentiation justifies re-considering not only the importance of one metric over the other, but also re-considering the entire metaphor. It is time for the experts to look at not only at adjusting TCP, but also at exploring other protocols, such as QUIC and others as well. It’s important that we feel free to consider alternatives to TCP. TCP is not a teddy bear, and one should not be afraid to replace it with a transport later with better properties benefiting users. A suggestion: we should consider desirable properties exercises. As we are looking at the parametric spaces, one can identify “desirable properties”, as opposed to “fundamental properties”. Among such properties, there may be a low-latency property. An example coming from ARPA: you want to know where the missile is now, not where it was. Understanding what is driving the particular parameter in the design space. When the parameter values are changed in extreme, such as connectiveness, some other designs will emerge. One case study is the interplanetary protocol, where “ping” is no long indicative of anything useful. While we look at responsiveness, we should not ignore connectivity. Unfortunately, maintaining backward compatibility is painful. The work on designing IPv6 so as to transition from IPv4 could have been done better if the backward compatibility was considered. This is too late for IPv6, but this problem space is not too late for the future laying problems. IPv6 is still not implemented fully everywhere. It’s been a long road since starting work in 1996, and we are still not there. In 1996, the thinking was that it was quite easy to implement IPv6, but that failed to hold true. In 1996 the dot-com boom started and lots of money was spent quickly, and the moment was not caught in time while the market expanded exponentially. This should serve as a cautionary tale. One last point: consider performance across multiple hops in the Internet. We’ve not seen many end-to-end metrics, as successfully developing end-to-end measurements across different network and business boundaries is quite hard to achieve. A good question to ask when developing new protocols is “will the new protocol work across multiple network hops?” Multi-hop networks are being gradually replaced by humongous flat networks with sufficient connectivity between operators so that systems become 1 hop or 2 hop at most away from each other (e.g. Google, Facebook, Amazon). The fundamental architecture of the Internet is changing.

The Internet is a shared network, built on the IP protocols using packet-switching to interconnect multiple autonomous networks. The Internet’s departure from circuit-switching technologies allowed it to scale beyond any other known network. On the other hand, the lack of in-network regulation made it difficult to ensure the best experience for every user. As the Internet use cases continue to expand, it becomes increasingly more difficult to predict which network characteristics correlate with better user experiences. Different application classes, e.g., video streaming and teleconferencing, can affect user experience in complex, and difficult to measure ways. Internet utilization shifts rapidly during the course of each day, week and year, which further complicates identifying key metrics capable of predicting a good user experience. Quality of Service (QoS) initiatives attempted to overcome these difficulties by strictly prioritizing different types of traffic. However, QoS metrics do not always correlate with user experience. The utility of the QoS metric is further limited by the difficulties in building solutions with the desired QoS characteristics. Quality of Experience (( QoE)) initiatives attempted to integrate the psychological aspects of how quality is perceived, and created statistical models designed to optimize the user experience. Despite these high modeling efforts, the QoE approach proved beneficial in certain application classes. Unfortunately, generalizing the models proved to be difficult, and the question of how different applications affect each other when sharing the same network remains open. The industry’s focus on giving the end-user more throughput/bandwidth led to remarkable advances. In many places around the world, a home user enjoys gigabit speeds to their Internet Service Provider. This is so remarkable that it would have been brushed off as science fiction a decade ago. However, the focus on increased capacity came at the expense of neglecting the other important core metric: latency. As a result, end-users whose experience is negatively affected by high lateness were advised to upgrade their equipment to get more throughput instead. showed that sometimes such an upgrade can lead to latency improvements, due to the economical reasons of overselling the “value-priced” data plans. As the industry continued to give the end user more throughput, while neglecting the latency metric, application designs started to employ various latency and short service disruption hiding techniques. For example, user experience of web browser performance is closely tired to the content in the local cache. While such techniques can clearly improve the user experience when using stale data is acceptable, this development is further decoupling user experience from the core metrics. In the most recent 10 years, efforts by Dave Taht and the bufferbloat society had led to significant progress updating queuing algorithms to reduce latencies under load compared to simipler FIFO queues. Unfortunately, the home router industry has yet to implement these algorithms, mostly due to marketing and cost reasons. Most home router manufacturers depend on System on a Chip (SoC) acceleration to to make products with a desired throughput. The SoC manufacturers opt for simpler algorithms and aggressive aggregation, reasoning that a higher-throughput chip will have guaranteed demand. Because consumers are offered choices primarily between different high throughput devices, the perception that a higher throughput leads to higher a quality of service continues to strengthen. The home router is not the only place that can benefit from clearer indications of acceptable performance for users. Since users perceive the Internet via the lens of applications, its important to appeal to the application vendors that they should adopt solutions that stress lower latencies. Unfortunately, while bandwidth is straightforward to measure, responsiveness is trickier. Many applications have found a set of metrics which are helpful to their realm, but these are not generalizable and universally applicable. Furthermore, due to the highly competitive application space, vendors may have economic reasons to avoid sharing their most useful metrics.

Measuring bandwidth is necessary, but is not alone sufficient. In many cases, Internet users don’t need more bandwidth, but rather need “better bandwidth” – i.e., they need other connectivity improvements. The users perceive the quality of their Internet connection based on the applications they use, which are affected by a combination of factors. There’s little value in exposing a typical user to the entire spectrum of possible reasons for the poor performance perceived in their application-centric view. Many factors affecting user experience are outside the users’ sphere of control. It’s unclear whether exposing the users to these other factors will help user’s understand their performance state. In general, users prefer simple, categorical choices (e.g. “good”, “better”, and “best” options). The Internet content market is highly competitive, and many applications develop their own “secret sauce.”

The workshop continued to discuss various metrics that can be used instead of or in addition to available bandwidth. Several workshop attendees presented deep-dive studies on measurement methodology.

Losing Internet access is, of course, the worst user experience. Unfortunately, unless rebooting the home router restores connectivity, there is little a user can do other than contacting their service provider. Nevertheless, there is value in the systematic collection of availability metrics on the client side: these can help the user’s ISP localize and resolve issues faster, while enabling users to better choose between ISPs. One can measure the availability directly by simply attempting connections from the client-side to locations of interest. For example, uses a large number of Android devices to measure network and cellular availability around the globe. Ookla collects hundreds of millions of data points per day, and uses these for accurate availability reporting. An alternative approach is to derive availability from the failure rates of other tests. For example, uses thousands of off-the shelf routers, called “Whiteboxes”, with measurement software developed by . These Whiteboxes perform an array of network tests and report availability based whether test connections were successful or not. Measuring available capacity can be helpful to the end-users, but it is even more valuable for service providers and application developers. High-definition video streaming requires significantly more capacity than any other type of traffic. At the time of the workshop, video traffic constituted 90% of overall Internet traffic and contributed to 95% of the revenues from monetization (via subscriptions, fees, or ads). As a result, video streaming services, such as Netflix, need to continuously cope with rapid changes in the available capacity. The ability to measure available capacity in real-time allows leveraging the different adaptive bitrate (ABR) compression algorithms to ensure the best possible user experience. Measuring the aggregated capacity demand allows Internet Service Provider’s to be ready for traffic spikes. For example, during the end-of-year holiday season, the global demand for capacity has been shown to be 5-7 times higher than other seasons. For end-users, knowledge of their capacity needs can help them choose a data plan best suited for them. In many cases, however, end-users have more than enough capacity, and adding more bandwidth will not improve their experience as after a point it is no longer the limiting factor in user experience. Finally, the ability to differentiate between the “throughput” and the “goodput” can be helpful in identifying when the network is saturated. In measuring network quality, latency is the time that it takes a network packet to traverse the path from one end to the other through the network. At the time of this report, users in many places worldwide can enjoy Internet access that has adequately high capacity and availability for their current needs. For these users, latency improvements, rather than bandwidth improvements, can lead to the most significant improvements in the quality of experience. The established latency metric is a round-trip time (RTT), commonly measured in milliseconds. However, users often find the RTT unintuitive since, unlike other performance metrics, high RTT values indicate poor latency. and presented an inverse metric, called “Round-trips per minute” (RPM). There is an essential distinction between the “idle latency” and “latency under working conditions.” The former is measured when the network is not used and reflects the best-case scenario. The latter is measured when the network is under a typical workload. Until recently, the typical case was to present the idle latency. However, these numbers can be misleading. For example, data presented at the workshop shows that the idle latency can be up to 25 times lower than the latency under typical working conditions. Because of that, when presenting latency to the end-user, it is essential to make a clear distinction between the two. Data shows that rapid changes in capacity affect latency. attempts to quantify how often a rapid change in capacity can cause connectivity to become “unstable”, i.e., having high latency but very little throughput. Such changes in capacity can be caused by infrastructure failures, but are much more often caused by in-network phenomena, such changing traffic engineering policies, or rapid changes in cross-traffic. Data presented at the workshop shows that 36% of measured lines have capacity metrics that vary by more than 10% throughout the day and across multiple days. These differences are caused by many variables, including local connectivity methods (WiFi vs. Ethernet), competing LAN traffic, device load/configuration, time of day and local loop/backhaul capacity. These factors make measuring capacity only using an end-user device or network difficult. A network router that sees aggregated traffic from multiple devices provides a better vantage point for capacity measurements. Such a test can account for the totality of local traffic and perform an independent capacity test. And even then, various factors might limit the accuracy of said test. Accurate capacity measurement requires a multiple samples. As users perceive the Internet through the lens of applications, it may be difficult to correlate changes in capacity and latency with the quality of the end-user experience. For example, web browsers rely on cached page versions to shorten page load times and mitigate connectivity losses. In addition, social networking applications often rely on pre-fetching their “feed” items. These techniques make the core in-network metrics less indicative of the users’ experience and necessitates collecting data in-application. It is helpful to distinguish between applications that operate on a “fixed latency budget” from those that have more tolerance to latency variance. Cloud gaming serves as an example application that requires a “fixed latency budget”, as a sudden latency spike can decide the “win/lose” ratio for a player. Companies that compete in the lucrative cloud gaming market make significant infrastructure investments, such as buiding entire datacenters closer to their users. These data centers highlight the economic benefits that having fewer latency spikes outweigh the associated deployment cost. On the other hand, applications that are more tolerant to latency spikes can sometimes operate reasonably well through short spikes. Yet even those applications can benefit from consistently low latency. For example, Video-on-Demand (VOD) apps can work reasonably well when the video is consumed linearly, but once the user tries to “switch a channel”, or to “skip ahead”, the user experience suffers unless the latency is sufficiently low. Finally, as the applications continue to evolve, in-application metrics are gaining in importance. Using VOD as an example, one can assess the quality of experience by checking whether the video player can use the highest possible resolution, whether the video is smooth or freezing, and other similar metrics. Then, the application developer can effectively use these metrics to prioritize future work. All popular video platforms (Youtube, Instagram, Netflix, and others) have developed frameworks to collect and analyze such metrics at scale. One example is the Scuba framework used by Meta . Unfortunately, the in-application metrics can be challenging to use for comparative research purposes. Firstly, different applications often use different metrics to measure the same phenomena. For example, application A can measure the smoothness of video via “mean time to re-buffer.” In contrast, application B can rely on the “probability of re-buffering per second” for the same purpose. A different challenge with using in-application metrics is that at the time of the workshop, VOD is a significant source of revenue for companies such as YouTube, Facebook, and Netflix, which places proprietary incentives against exchanging the in-application data. Finally, in-application metrics can also accurately describe the activities and preferences of an individual end-user, leading to privacy infringements.

Availability is simply defined as whether or not a packet can be sent and then received by its intended recipient. Availability is naively thought to be the simplest to measure, but is more complex when considering that continual, instantaneous measurements would be needed to detect the smallest of outages. Also difficult is determining the root cause of infallibility: was the user’s line down, something in the middle of the network or was it the service with which the user was attempting to communicate.

If the network capacity does not meet the user demands, the network quality will be impacted. Once the capacity meets the demands, increasing capacity won’t lead to further quality improvements. The actual network connection capacity is determined by the equipment and the lines along the network path, and it varies throughout the day and across multiple days. Studies involving DSL lines in North America indicate that over 30% of the DSL lines have capacity metrics that vary by more than 10% throughout the day and accross multiple days. Some factors that affect the actual capacity are: Presence of a competing traffic, either in the LAN or in the WAN environments. In the LAN setting, the competing traffic reflects the multiple devices that share the Internet connection. In the WAN setting the competing traffic often originates from the unrelated network flows that happen to share the same network path. Capabilities of the equipment along the path of the network connection, including the data transfer rate and the amount of memory used for buffering. Active traffic management measures, such as traffic shapers and policers that are often used by the network providers. There are other factors that can negatively affect the actual line capacities. The user demands of the traffic follow the usage patterns and preferences of the particular users. For example, large data transfers can use any available capacity, while the media streaming applicaitons require limited capacity to function correclty. Video-conferencing applications typically need less capacity than high-definition video streaming.

End-to-end latency is the time that a particular packet takes to traverse the network path from the user to their destination and back. The end-to-end latency comprises several components: The propagation delay, which reflects the path distance and the individual link technologies (e.g. fibre vs satellite). The propagation doesn’t depend on the utilization of the network, to the extent that the network path remains constant. The buffering delay, which reflects the time segments spend in the memory of the network equipment that connect the individual network links, as well as in the memory of the transmitting endpoint. The buffering delay depends on the network utilization, as well as on the algorithms that govern the queued segments. The transport protocol delays, which reflects the time spent in retransmission and reassembly, as well as the time spent when the transport is “head-of-line blocked.” Some of the workshop sumbissions have explicitly called out the application delay, which reflects the inefficiencies in the application layer. Traditionally, end-to-end latency is measured when the network is idle. Results of such measurements reflect mostly the propagation delay, but not other kinds of delay. This report uses the term “idle latency” to refer to results achieved under idle network conditions. Alternatively, if the latency is measured when the network is under its typical working conditions, the results reflect multiple types of delays. This report uses the term “working latency” to refer to such results. Other sources use the term “latency under load” (LUL) as a synonym. Data presented at the workshop reveals a substantial difference between the idle latency and the working latency. Depending on the traffic direciton and the technology type, the working latency is between 6 to 25 times higher than the idle latency: Direction Technology type Working latency Idle latency Working - Idle difference Working / Idle ratio Downstream FTTH 148 10 138 15 Dowstream Cable 103 13 90 8 Downstream DSL 194 10 184 19 Upstream FTTH 207 12 195 17 Upstream Cable 176 27 149 6 Upstream DSL 686 27 659 25 While historically the tooling available for measuring latency focused on measuring the idle latency, there is a trend in the industry to start measuring the working latency as well, e.g. .

The participants have proposed several concrete methodologies for measuring the network quality for the end users. introduced a methodology for measuring working latency from the end-user vantage point. The suggested method incrementally adds network flows between the user device and a server endpoint until a bottleneck capacity is reached. From these measurements, a round trip latency is measured and reported to the end-user. The authors chose to report results with the RPM metric. The methodology had been implemented in Apple Monterey OS. have applied the RPM metric to the results of more than 4 billion download tests that M-Lab performed in 2010-2021. During this time frame, the M-Lab measurement platform underwent several upgrades which allowed the research team to compare the effect of different TCP congestion control algorithms (CCAs) on the measured end-to-end latency. The study showed that the use Cubic CCA leads to increased working latency, which is attributed to its use of larger queues. presented a large-scale study that aimed to establish a correlation between goodput and quality of experience on a large social network. The authors performed the measurements at multiple data centers from which video segments of set sizes were streamed to a large number of end users. The authors used the goodput and throughput metrics to determine whether particular paths were congested. presented the analysis of working latency measurements collected as part of the FCC’s “Measuring Broadband America” (MBA) program. The FCC does not include working latency in its yearly report, but does offer it in the raw data files. The authors used a subset of the raw data to identify important differences in the working latencies across different ISPs. presented analysis of working latency across multiple service tiers. They found that, unsurprisingly, “premium” tier users experienced lower working latency compared to a “value” tier. The data demonstrated that working latency varies significantly within each tier; one possible explanation is the difference in equipment deployed in the homes. These studies have stressed the importance of measurement of the working latency. At the time of this report, many home router manufacturers relied on hardware-accelerated routing which used FIFO queues. Focusing the working latency measurements on those devices, and making the consumer aware of the effect of chosing one manufacturer vs. other can help improving the home router situation. The ideal test would be able to identify the working latency, and to pinpoint to the source of delay (home router, ISP, server side, or some network node in between). Another source of high working latency comes from network routers that are exposed to cross-traffic. As indicated, these can become saturated during the peak hours of the day. Systematic testing of the working latency in routers under load can help improve the infrastructure.

The metrics for network quality can be roughly grouped into: Availability metrics, which indicate whether the user can access the network at all. Capacity metrics, which indicate whether the actual line capacity is sufficient to meet the user’s demands. Latency metrics, indicating if the user gets the data in a timely fashion. Higher-order metrics, which include both the network metrics, such as inter-packet arrival time, and the applicaiton metrics, such as the mean time between rebuffering for video streaming. The availabiltiy metrics can be seen as derivative of either the capacity (zero capacity leading to zero availability) or the latency (infinite latency leading to zero availability). Key points from the presentations and discussions included: Availability and capacity are “hygienic factors” - unless an application is capable of using extra capacity, end-users will see little benefit from using overprovisioned lines. The working latency has stronger correlation with user experience than latency under an idle network load. The working latency can exceed the idle latency by order of magnitude. The RPM metric is a stable metric, with positive values being better, that can be effective to communicate latency to the end-users. The relationship between throughput and goodput can be effective in finding the saturation points, both in client-side and server-side settings. Working latency depends on algorithm choice for addressing endpoint congestion control and router queuing. Finally, it was commonly agreed to that the best metrics are those that are actionable.

In the Cross-layer section participants presented material and discussed how to accurately measure exactly where problems occur. The discussion showed how difficult it is to achieve accuracy when many components of a network connection affects the measurements. Discussion centered especially on the differences between physically wired and wireless connections and the difficulties of accurately determining problem spots when multiple different network types are responsible for the quality. As an example, showed that as Internet access becomes the norm, the limited bandwidth of 2.4Ghz wifi is most frequently the bottleneck. In comparison, the wider bandwidth of the 5Ghz WiFi have only been the bottleneck in 20% of the observations. The participants agreed that no single component of a network connection has all the data required to measure the effects of the network performance on the quality of the end user experience. The applications that are running on the end-user devices have the best insight into their respective performance, but have limited visibility into the behavior of the network, and are not able to act on the limited information about the network performance. Internet service providers have good insight into QoS considerations, but are not able to infer the effect of the QoS metrics on the quality of end user experiences. Content providers have good insight into the aggregated behavior of the end users, but lack the insight on what aspects of the network performance are leading indicators of user behavior. The workshop had identified the need for a standard and extensible way to exchange network performance characteristics. Such an exchange standard should address (at least) the following: A scalable way to capture the performance of multiple (potentially thousands of) endpoints. The need for an accompanying set of tools to analyze the data. A transparent model for giving the different actors on the network connection an incentive to share the performance data they collect. Preservation of end-user privacy. In particular, federated learning approaches, where no centralized entity has the access to the whole picture, should be preferred. The data exchange format should include precautions against data manipulations, so that the different actors won’t be tempted to game the mechanism.

Commonly, there’s a tight coupling between collecting performance metrics, interpreting those metrics and and acting upon the intrepretation of the metrics. Unfortunately, such model is not the best for successfully exchanging cross-layer data: The actors that have the ability to collect particular performance metrics (e.g. the TCP RTT) do not necessarily have the context necessary for a meaningful interpretation. The actors that have the context and the computational/storage capacity for the interpretation do not necessarily have the abilty to control the behavior of network / application. The actors that can control the behavior of network / application typically do not have access to the data. The participants agreed that it is important to separate the above three aspects, so that: The different actors that have the data but not the ability to interpret / act upon should publish their measured data. The actors that have the expertise in interpreting and synthesizing the performance data will be able to publish the results of any interpretation.

Preserving the privacy of the end users is a difficult requirement to meet when addressing this problem space. There is an intrinsic trade-off between collecting more data about user activities, and infringing their privacy in doing so. Participants agreed that observability across multiple layers is necessary for an accurate measurement of the network quality.

The TCP protocol makes several metrics available for passive measurement, and the following metrics have been found to be effective: TCP connection latency measured using SACK/ACK timing, as well as the timing between TCP retransmission events, are good proxies for end-to-end RTT measurements. On the Linux platform, the tcp_info structure is the de-facto standard for an application to introspect the performance of kernel-space networking. However, there is no equivalent de-facto standard for the user-space networking. The QUIC and MASQUE protocols make passive performance measurements more challenging. An approach that uses federated measurement / hierarchical aggregation appears more valuable for these protocols. The QLOG format seems to be the most mature candidate for such an exchange.

The ownership of the Internet is spread across multiple administrative domains, making measuring performance data difficult. Furthermore, the immense scale of the Internet makes aggregation and analysis of such data difficult. presented a simple logging format that could potentially be used to collect and aggregate data from different layers. Another aspect of cross-layer collaboration hampering measurement is that the majority of current algorithms do not explicitly provide performance data that can be used in cross-layer analysis. The IETF community can be more diligent in identifying a protocol’s key performance indicators, and exposing those as part of the protocol specification. Despite all the challenges, it should still be possible to perform limited-scope studies in order to have a better understanding of how user quality is affected by the interaction of the different components that constitute the Internet. Recent development of federated learning algorithms suggests that it might be possible to perform cross-layer performance measurements while preserving user privacy.

With the advent of the L4S congestion notification and control, there is an even higher need for the transport protocols and the underlying hardware to work in unison. At the time of the workshop, the typical home router used a single FIFO queue, large enough to allow amortizing the lower-layer header overhead across multiple transport PDUs. These designs worked well with the Cubic congestion control algorithm, yet the newer generation of CCAs can operate on much smaller queues. To fully support latencies less than 1ms, the home router needs to work efficiently on sequential transmissions of just a few segments vs. being optimized for large packet bursts. Another design trait that’s common in home routers is the use of packet aggregation to further amortize the overhead added by the lower-layer headers. Specifically, multiple IP datagrams are combined into a single large tranfer frame. However, this aggregation can add up to 10ms to the packet sojourn delay. Following the famous “you can’t improve what you don’t measure” adage, it is important to expose these aggregation delays in a way that would allow identifying the source of the bottlenecks, and making hardware more suitable for the next generation transport protocols.

TBD

Finally, in the Synthesis section presentations and discussions concentrated on the next steps likely needed to make forward progress. Of particular concern is how to bring forward measurements that can make sense to end users trying to make subscription decisions. (this section is TBD)

(this section is TBD)