Packet loss is an inevitable fact of life and affects the performance of all latency-sensitive applications. Its impact on the perceived “speed” of the applications and hence on their quality of experience (QoE) is most noticeable by mobile users.
Packet loss averages 1.5-2.5% per user on wireless networks, but has been measured to reach 5% and higher. More importantly, with 2% packet loss, users can lose 97% of their TCP throughput (LTE profile).
And to make things worse, because packet loss is transient, there is no way to know which users will be affected at any point in time, or its frequency. The only way to adequately mitigate packet loss is to address it in real-time with AI-powered, (non-rules-based) adaptive systems.
Packet loss is unpredictable and transient.
Packet loss has a direct impact on revenue for eCommerce businesses. Amazon has reported a 1% increase in revenue when latency is reduced by 100ms. At scale, the impact is quite large. Amazon’s 2017 US eCommerce sales were roughly $200 billion, so a 1% improvement would yield an additional $2 billion in revenue.
- 1% revenue increase per 100ms latency reduction
- 28% of viewers abandon streaming video after just one rebuffering event
- 51% of viewers stop watching video after 2 or 3 buffer delays
Online businesses live and die based on the responsiveness of, and hence the engagement with their application. Poor throughput, a direct result of packet loss, detracts from the user’s Quality of Experience (QoE) and reduces the time she spends with the application.
Performance of online video-on-demand and live streaming events suffers in the face of packet loss as well. A study among OTT viewers has shown that this group is fickle in their attachment to an OTT video service, with one in three abandoning the current provider as soon as they have an inferior experience. And 51% of viewers of streaming video say they stop watching after 2 or 3 attempts when they encounter buffer delays, while 28% stop after encountering just one interruption.
Our own studies show that large numbers of wireless users in the U.S. experience packet loss at any moment in time. This has detrimental effects on users while shopping, watching streaming video, playing online games, and performing other latency-sensitive activities online.
During a recent major media video streaming event we tracked roughly 50 million packets to show the amount of packet loss affecting real world users.
Of the connections observed during the streaming event, roughly 90% percent of total connections experienced packet loss. The median packet loss percentage was 5.48% and the average packet loss was 12.62%, shattering industry averages by over 3X.
- This graph shows packet loss by individual user. Each bar represents one user. Different colors correspond to different ISPs.
- Y-axis: Percentage of packet loss experienced by the viewer.
- X-axis: Individual viewer IP addresses.
- Time period: 1 hour.
- Total # of packets tracked in this graph:
- Approximately 50 million.
- Geographical location: US and Canada traffic.
- During the streaming event, 90% of all connections experienced packet loss
- Median packet loss: 5.48%
- Average packet loss: 12.62%
Traditional delivery networks or edge platforms do not mitigate or even measure packet loss. So without Zycada, users on networks with packet loss suffered from reduced throughput from the origin, leading to increased video start times, lower average bitrates, and more buffer-empty events.
Zycada has measured the prevalence of packet loss for over 120 billion transactions of real-world traffic and how it impacts throughput — perceptible as a reduction in overall speed.
|Rated Bandwidth||Packet Loss||Actual Throughput|
|90 Mbps||0%||90 Mbps|
|90 Mbps||2%||3 Mbps|
Throughput is decreased by 97%! Make no mistake: The user experience is severely compromised with 2% packet loss. Slow checkouts in eCommerce stores and more buffer-empty events in video are guaranteed to occur.
Packet loss in the wireless last mile makes delivering a great user experience nearly impossible. Putting content in Points of Presence (PoPs) closer to the user only reduces round trip times based on physical distance. But caching capacity at the edge is finite and subject to frequent evictions. And it does nothing to mitigate lost packets.
The Zycada Delivery Network is built from the ground up to tackle the hardest performance problems faced in the wireless last mile. It is powered by AI and automatically mitigates packet loss for each connection. Instead of optimizing the server for throughput, Zycada fingerprints and individually optimizes each connection between the client and the origin. Zycada’s client-aware approach provides the opportunity to address last mile performance problems such as packet loss, packet corruption and jitter. It requires no code changes or SDK on the client or the origin.
- Zycada mitigates last mile performance problems.
- It requires no SDK or code changes on client or origin.
As last mile performance problems are unpredictable and transient, they cannot be modeled; they happen live, so any solution addressing them has to be able to act in real-time. As content flows from the delivery network to the user, Zycada’s AI algorithms continuously adapt to that specific connection and optimize the packet flow in real-time. This ensures a consistently smooth delivery for all types of content.
Generated by a major Zycada eCommerce customer due to reduced latency
Most users consume the Internet over wireless media. Traditionally, DevOps teams optimize throughput at the server level in an attempt to maintain high performance for wireless clients. But the improvements stop at the server — they cannot optimize the individual connection from the client to the server, nor the performance on the wireless last mile. This results in a “one-size-fits-all” approach to dealing with network problems – every client, regardless of its connection characteristics, receives the same treatment.
Zycada’s approach is to look at the connection from the client’s view out to the origin and orient the delivery network’s perspective around that client. For each client, packet loss mitigation, dynamic routing, and origin health scores are monitored in real-time. Packet loss mitigation is one such benefit of the connection profile, but in reality, packet loss, origin health and load all contribute to latency. Zycada’s solution takes all these variables into account and constructs a unique flow profile (i.e. shaping of packet flow) to maximize the throughput for a given transaction. And Zycada works with any existing CDN architecture – single CDN, multi-CDN and multi-cloud approaches are all supported.
How Zycada Fights Packet Loss
|With Zycada Delivery Network||Without Zycada Delivery Network|
|Applies packet loss mitigation automatically.||Does not mitigate packet loss.|
|Every connection to the Zycada delivery network is characterized and receives its own packet loss mitigation optimizations. These are updated in real-time. Each end device has a uniquely identifiable flow profile.||As a one-size-fits-all approach, any load time optimizations are applied to all connections in the same way, at the CDN server level.|
|All measurements occur in real-time (in-band). App availability does not affect the delivery of metrics. If the client app crashes, measurements are still ongoing.||Measurements occur after content arrives (out of band). Telemetry relies on the app to be available. If the client app crashes, measurements will not be delivered.|
|Zycada Delivery Network's optimizations span sockets, sessions, and domains from client to origin.||Can apply improvements only within the context of a socket.|