Negative latency in the banker's kitchen
This technique was explained to me when I went to visit the Saskatoon office of my current employer. Here's how it works. When you make tea, you need to boil water, then let it cool down a bit. But if you're lucky, there was hot water ("response") already in the kettle, so you can use that. If not, you make your own, then use that. Afterward - and here's the magic - you refill the kettle and turn it on (submit a "prequest") for the next person, thus greatly increasing the chance that the next person ("request") to arrive will already have hot water.
This only works in a situation with sufficiently high tea throughput and sufficiently little clustering of requests, and you sometimes end up requesting extra data (boiling extra water) at the end of the day, but the bandwidth cost of some boiled water is much less than the latency cost of waiting for water to boil.