Preparation

Load testing builds confidence in your infrastructure design before a large event and highlights weaknesses early. When possible, use historical data to model and test system load. For example, if you’ve acquired content streaming rights, but a similar show or event has aired in the past, Nielsen or Comscore data can provide useful estimates for peak concurrent viewers. With this model, you can develop accurate load tests that simulate real-world conditions. Be sure to test both individual service components and the end-to-end system with load tests that emulate end user clients.

For origin services, estimate load in Transactions Per Second (TPS) or Requests Per Second (RPS). We advise that you estimate peak TPS the player devices will generate based on the usage pattern, layers of caching, and unique nature of your audience. 

For example, for an HLS stream with 6-second segments and un-muxed video, audio, and captions, each player will generate one TPS (= 3*1/6 TPS for video, audio, and caption manifest requests + 3*1/6 for video, audio, and caption segment requests) to the CDN. Assuming a 95% cache hit ratio (CHR) for the CDN, it will result in 0.05 TPS (=1*(1-0.95)) to the origin. To calculate peak TPS to the origin, multiply the peak number of player devices expected at any time with 0.05 TPS.

For DASH streams, the TPS depends on how frequently the DASH player requests the manifest and segments. A common practice is to request once every segment length. Note, that all DASH segments are unmuxed, that is, the player will request a video, audio, and caption segments each time. So, for a 6-second segment, the TPS would be 0.66 TPS (= 1/6 for manifest + 3*1/6 for video, audio and caption segments). Again, assuming a 95% cache hit ratio (CHR) for the CDN, that would result in 0.033 TPS (=0.66*(1-0.95)) to the origin. Based on the forecast for peak concurrent HLS and DASH playback sessions, peak TPS for the origin can then be calculated.

You can use the following TPS formulas to guide your estimations, but, when available, always use your own metrics and data to forecast usage.

Estimating transactions per second for common streaming media protocols

For HLS, the number of elementary stream segments depends on whether the video, audio, and caption date is muxed in one segment or unmuxed in separate segments. In case of muxed segments, there is only one elementary stream segment whereas in the case of unmuxed segments, there are three elementary stream segments.

AWS Cloud can handle very large scale streaming events. However, services have soft limits in place to protect customers from inadvertently scaling resources beyond their needs and racking up unnecessary costs. Service quotas (also referred to as limits) for all media services are Region-specific unless otherwise noted in the documentation. For more information about the quotas for a specific service, refer to its documentation. 

You can centrally manage Service Quotas for all services in your account using Service Quotas in the AWS Management Console where you can view your current quotas and also request increases. If you don’t find your AWS service in Service Quotas, contact AWS Support to increase them. Provide details of the Regions in which your workflow will operate, usage patterns, and timeframe.

Consider using Infrastructure and Event Management (IEM) and Media Events Management programs for large-scale live events that might require immediate support during an event. By engaging AWS through these programs, you enable AWS experts to become familiar with your workload, provide architectural and operational guidance, and real-time support for your planned event.

In addition to IEM, Enterprise Support customers are eligible for a Cloud Operations Review and a Well-Architected Framework Review designed to help identify risks in your cloud operations. This cross-team engagement helps establish a common understanding of your workload and helps AWS contribute to your streaming events. You can always perform your own architecture review for any workload by using the AWS Well-Architected Tool in the AWS Management Console.