Wesley Chan

I write the units into every figure, because a transfer estimate without them is a number rather than an answer. The subtlety is which bytes. A file on disk is not the number crossing the network, and the difference is where the surprise lives. Protocol overhead is the first layer. A header accompanies each segment, and acknowledgement segments travel back the other way with their own headers. That is why a transfer of many small segments is slower than one of the same total size in large segments, and why packet count matters as much as byte count. Compression is the second, and it changes the arithmetic rather than adding to it. Compressing before sending reduces the bytes that cross the wire, so a file that compresses to a fifth of its size can transfer faster than one of the same size that does not. That is also why compressing something already compressed costs time and saves nothing. Rate units are the third. A figure quoted in megabits and a figure quoted in megabytes differ by eight, and the mistake is common enough that I lead with it. Transfer time is a steady-state estimate, so I cover what it excludes: the setup, the acknowledgement pattern, anything else competing for the link, and any step where the sending side waits rather than sends. I finish each page with the same advice, which is to measure the real transfer rather than model it. Every input to this arithmetic is uncertain, and the uncertainty compounds.