Comparisons and explanations · Intermediate

Bandwidth versus latency

Distinguish transfer capacity from delay using serialization time, round trips, queueing and parallel-request examples.

Ask Logeia about this articleTry the three questionsReview your answers

Updated

On this page
  1. Do not collapse performance into one number
  2. Calculate the serialization component
  3. Round trips can dominate small dependent requests
  4. Queueing changes delay under load
  5. Interpret the bandwidth-delay product carefully
  6. Compare equivalent completed work
  7. Ask Logeia about this articleCheck your understanding

Do not collapse performance into one number

Bandwidth concerns data-carrying capacity over time, whereas latency concerns delay before an observed result arrives. A reported latency may be one-way delay, round-trip time, or the complete duration of an API operation; the measurement definition decides which. A larger Mbps figure does not necessarily make a tiny interactive request faster. Compare the number of bytes, the path, and the start and completion events. Network capacity and application response duration are not interchangeable columns in a performance table. Establishing these units first prevents an expensive capacity upgrade from being presented as a remedy for a delay caused somewhere else in the request.

Calculate the serialization component

For an illustrative calculation, send one million bytes at one hundred million bits per second, ignoring headers, loss, and startup. Eight million bits require 0.08 seconds of serialization. Propagation, queueing, and server work add other delay components. Ten times the capacity reduces this particular transfer component to 0.008 seconds; it does not divide every fixed waiting component by ten. Mixing bits with bytes or decimal MB with binary MiB changes the arithmetic. These values explain a model rather than report a measured connection. State omitted costs before comparing its result with a stopwatch measurement of a real protocol exchange.

Round trips can dominate small dependent requests

Imagine ten API calls where each request depends on the previous response. If each round trip contributes eighty milliseconds and transfer and processing costs are ignored, the waits alone total about eight hundred milliseconds. More bandwidth does not remove those ten dependency-imposed exchanges. Batching information, issuing independent requests concurrently, or reusing connections can address different sources of overhead. Connection reuse, however, does not by itself eliminate ten application-level dependencies. Explain which waits a proposed change removes. Otherwise improvements from fewer round trips may incorrectly be attributed to a faster transfer rate, leading to the wrong design on a different workload.

Queueing changes delay under load

Even on the same link, bursts can increase latency by creating a queue before transmission. Fully utilizing capacity is not the same as providing good response time to each user. Enlarging a queue without controlling offered load may reduce drops while allowing requests to wait longer. Inspect slower percentiles and timeout counts, not just the average among completed operations. Counting only successes can omit the users who experienced the worst outcome. Align offered work and delay on the same timeline. This helps distinguish a steady propagation cost from a load-sensitive queueing problem and shows whether an intervention improved capacity, waiting, or merely the accounting.

Interpret the bandwidth-delay product carefully

The bandwidth-delay product estimates data in flight associated with a sustained transfer model. Using one hundred Mbps and an eighty-millisecond round trip gives eight million bits, approximately one million bytes. This is neither an application message size nor unconditional permission to transmit that much data safely. TCP receive windows, congestion control, loss, and buffering also matter. The product helps explain why a high-capacity, long-round-trip path may be underutilized when too little data can remain in flight. Specify whether one-way or round-trip delay is being used and the purpose of the calculation; the numerical product alone does not diagnose which limiting mechanism is active.

Compare equivalent completed work

Match input size, connection state, concurrency, path, and cache conditions. A new connection including TLS setup is not directly comparable to body transfer on an established reused connection. Include both small requests and larger transfers to distinguish fixed waiting from size-dependent costs. Measure correctly completed work as well as bytes sent; dropped or timed-out work can make a throughput figure misleading. If greater link capacity does not change perceived responsiveness, examine serial application operations and external-service waits before declaring the capacity measurement wrong. The outcome may simply show that transfer time was a small fraction of the user’s original end-to-end delay.

Reading and practice are recorded separately in this browser, for each language.

Learn this in order

Understand the systems behind the web · STEP 10

Choose a learning path