About Latency Simulator

Simulate latency impact on performance with waterfall analysis and cascading delays. Part of the DevTools Surf developer suite. Browse more tools in the Developer Utilities collection.

Use Cases

• Model a multi-service request chain to identify which downstream dependency contributes most to total latency.
• Simulate the latency impact of moving from a monolith to microservices with additional network hops.
• Estimate P99 latency for a system under load before deploying to production.
• Compare parallel vs. sequential API call strategies to justify an architecture decision.

Tips

• Model dependent API calls as sequential, not parallel, to see worst-case latency from chained requests.
• Add jitter to each step to simulate realistic network variance — production latency is never a fixed number.
• Compare P50 vs. P99 latency in the waterfall — tail latencies (P99) determine user experience for the slowest 1% of requests, which are often the most frustrated users.

Fun Facts

• Amazon's 2006 internal research found that every 100ms of latency cost them 1% in sales — one of the first quantified business impact measurements of web performance, cited in countless engineering talks since.
• The speed of light limits transatlantic fiber latency to ~70ms round-trip at minimum. Real RTT is 80-100ms due to routing overhead — physics creates an irreducible floor regardless of server optimization.
• A 2014 Google study on mobile search found that adding 400ms of latency reduced search volume by 0.44% — users abandon slower pages even when they don't consciously notice the delay.

FAQ

What's a waterfall analysis?

A sequential timeline showing each request or processing step, its start time, and duration. The total request time is visible as the sum of cascading steps — identical to browser DevTools Network waterfall.

How does it model P95/P99 latency?

Each step gets a distribution (mean + standard deviation). Monte Carlo simulation runs 10,000 iterations, sampling from each distribution and aggregating end-to-end latency to produce percentile outputs.

Can I model parallel requests?

Yes — group steps as 'parallel' and the simulator takes the max of the group as the effective duration, then adds the coordination overhead. Useful for fan-out/fan-in patterns.