Carnegie Mellon Algorithm Cuts Data Center Waste by Half

For three decades, computer scientists have watched helplessly as one of their field's most fundamental problems stayed stubbornly hard. They wondered how to distribute tasks evenly across two groups when they don't know in advance what the tasks will weigh.

It sounds abstract until you realize it describes everything from hospital operating theaters to data centers to power grids. Last month, researchers at Carnegie Mellon and Princeton found a better algorithm.

What the breakthrough papers won't tell you is what was actually constrained while this problem stayed unsolved. Cloud providers engineered around imbalance by overprovisioning capacity by 30 to 50 percent—keeping spare servers spinning so that when traffic lurches toward one side of the network, nothing breaks.

This is the technological equivalent of a safety margin you pay for every day whether you need it or not. Data centers consume enormous amounts of electricity. That cost distributes backward through every subscription, every API call, every startup trying to scale on a limited budget. Better allocation means less wasted capacity, which means lower costs for cloud providers and lower prices for customers.

A $50-a-month subscription to a managed database might drop to $37. That difference is almost invisible at scale until you're a three-person company that was going to build its own system because buying the service was unaffordable—suddenly it's not. The algorithm solved imbalance. The economy will parse it as a door opening for people who were locked out before.

The algorithm solved imbalance. The economy will parse it as a door opening for people who were locked out before.