Latency is how long one round trip takes; bandwidth is how many bits a second the link carries once the answer is flowing. They are two numbers, and their product — the bandwidth-delay product — is how many bytes the wire holds at once. A transfer smaller than that finishes before the pipe is full, and for it the bandwidth figure is very nearly irrelevant.
Core ideas
One hard idea, and it is that there are two numbers here rather than one. Everything else on this page is the arithmetic of deciding which of the two you are paying.
The round-trip time RTT is a duration: how long a signal takes to reach the far end and an answer to come back. Bandwidth is a rate: how many bits a second the link carries once data is moving. Buying a bigger link raises the rate. It does not shorten the distance, so it does not move the duration, and the two can be changed independently of each other.
A rate multiplied by a time is a quantity. 100 Mbit/s × 80 ms is 1 MB: that much data can be on the wire at once, sent and not yet arrived. That is the bandwidth-delay product, and it is measured in bytes rather than in bytes a second, which is what makes it comparable with the size of a transfer.
Time is RTT + size/bandwidth. Well below the BDP the first term is nearly all of it and the transfer is latency-bound; well above it the second takes over and the transfer is bandwidth-bound. No property of the link settles this on its own. The size does, against the link.
The material
What a bandwidth-delay product is a quantity of, and how to read a transfer’s size against it.
The bandwidth-delay product of a link is its bandwidth multiplied by its round-trip time. A rate in bits a second times a time in seconds is a number of bits, so the product is a quantity of data: how much can be in flight — sent, and not yet acknowledged — at any instant.
It is worth doing that conversion once, slowly, because the units are where this goes wrong. 100 Mbit/s is 10⁸ bits a second, which is 1.25 × 10⁷ bytes a second. Eighty milliseconds of that is 10⁶ bytes — one megabyte, reading 1 kB as 1000 B, which is how link rates are quoted and how this course counts. So a 100 Mbit/s link with an 80 ms round trip has a megabyte of road on it.
A 64 kB transfer over that link costs one round trip of 80 ms plus 64 000 × 8 / (100 × 1000) = 5.12 ms of sending. That is 85.12 ms in all, of which 94.0% is waiting. The sending term is the smaller number by a factor of more than fifteen, and it is the only one bandwidth touches.
Double the bandwidth to 200 Mbit/s and the sending term halves to 2.56 ms: the transfer saves 2.560 ms and now takes 82.56 ms. Halve the round trip to 40 ms instead and it saves 40 ms, finishing in 45.12 ms. The second change is worth about sixteen times the first, on the same transfer, and no amount of further bandwidth ever closes that gap.
There is exactly one size at which the two terms cost the same: the size for which size/bandwidth = RTT, which rearranges to size = bandwidth × RTT. That is the bandwidth-delay product again. The BDP is not merely a capacity; it is the crossover, and a transfer is latency-bound or bandwidth-bound according to which side of it the transfer falls on.
4 MB over the same 100 Mbit/s link at 80 ms spends 320 ms sending and 80 ms waiting: 400 ms in all, of which the round trip is 20.0%. This transfer is past the 1 MB crossover and is bandwidth-bound. Here doubling the link saves 160 ms and halving the round trip saves 40 ms — the advice from the previous example, applied here, is four times worse than the alternative.
The misconception worth naming out loud is that a faster link lowers the round-trip time. It does not. The round trip is set by how far the signal goes and how many devices it passes through; a bigger pipe moves more per second through the same pipe, at the same distance. Someone whose page is slow because it makes four sequential requests over an 80 ms link can buy a gigabit connection and measure no change at all.
What does move the round trip is shortening the distance or removing a trip. “The Speed-of-Light Floor” puts a hard bound under the first of those, and “Round Trips, Not Bytes” counts the second, which is usually the one you can change.
Interactive
The bandwidth-delay product, and which term your transfer is paying
Method
Four lines of arithmetic, and the fourth is the one that stops you buying the wrong thing.
Divide the bit rate by eight. 100 Mbit/s is 1.25 × 10⁷ B/s. Doing this first is what prevents the factor-of-eight error, which is the most common arithmetic mistake in this whole subject.
Bytes a second times seconds is bytes. Keep the time in seconds for this one line even though the rest of the course is in milliseconds, then convert back; mixing the two here is how a thousand-fold error gets in.
Smaller means latency-bound: the answer is roughly one round trip and the bytes are a rounding error. Larger means bandwidth-bound: the answer is roughly size/bandwidth and the round trip is the rounding error.
Compute what halving the round trip saves and what doubling the bandwidth saves, in milliseconds, on this transfer. They are rarely close, and the larger one is almost never the one that was already being proposed.
The link never changes. Only the size does, and it changes the answer to “what should we buy?” completely.
The two verdicts point in opposite directions on the same link, which is the point. “Is this link fast?” has no answer. “Is this transfer fast on this link?” has one, and it is decided by the size against the bandwidth-delay product.
Notice what halving the round trip saves in both rows: 40 ms, exactly, regardless of the transfer size. The round-trip term does not know how big the payload is. That is why it is the term that dominates small transfers and the term that vanishes into the noise on large ones.
For a faded rehearsal, take a 20 Mbit/s link with a 200 ms round trip — a satellite hop. The supplied first move is the conversion: 20 Mbit/s is 2.5 × 10⁶ B/s. Compute the bandwidth-delay product, then decide what a 300 kB transfer is bound by, and how many milliseconds each of the two upgrades would save it. Write down which side of the crossover you are on before you open the quiz, and check the figure in the lab by setting those three sliders.
You should be able to convert a link rate to bytes a second, compute its bandwidth-delay product, classify a given transfer against that product, and state in milliseconds what each of the two possible upgrades would save on that transfer.
Something to change about this lesson? Write it down, tick it off when it is handled, and export the list when you want to act on it. Kept in this browser only: nothing is sent anywhere, and nobody is notified.
No recommendations yet for this lesson.