Cloud Servers Waste Memory — This Fix Could Cut Your App Bills by 40%

Every time you open an app, order food, or check your bank balance, your request lands on a rented computer in a data center. Companies rent these computers from Amazon, Google, or Microsoft. The problem: those rentals come in fixed bundles. You pick a memory size, and the computing power is locked to it. If your app needs lots of computing but little memory, you still pay for memory you never use, and the leftover capacity on each machine sits idle.

The researchers call this "fragmentation" — think of packing a moving truck with boxes of only two shapes. You end up with gaps you can't fill, so you need a second truck. Their fix, Memoryless, treats memory like a utility you can borrow from across the room instead of something bolted to your desk. Apps get a mix of fast local memory and slightly slower memory from a neighboring machine. The scheduler then picks the app version that best fits whatever gaps are left, like a Tetris player who reshapes the pieces to fill every row.

Built on Knative, a popular open-source tool for running apps without managing servers, Memoryless was tested with realistic traffic patterns. It used up to 40% fewer machines during normal operation and 46% fewer at peak times. Crucially, it kept service quality within 3% of normal — meaning users wouldn't notice slower apps. It also cut stranded computing power by 44%.

Why should you care? Cloud costs are baked into nearly every subscription you pay. When companies waste servers, that waste shows up in your monthly bills, in slower launches of new features, and in the energy data centers burn. If this approach spreads, the same hardware could serve more people, and companies could pass those savings along. The catch: borrowed memory travels over a network, so it's a bit slower than memory sitting right next to the processor — fine for most everyday apps, but not for the most demanding ones.