New Software Speeds Up AI Training When Cloud Networks Jam

Creating today's AI models takes groups of powerful computers working like a giant team. During training, every machine must constantly swap information with the others — think of a hundred people passing notes to solve one puzzle together. On shared cloud servers, another company's AI traffic can clog one of those connections, and the entire team waits. This slowdown is a huge headache for businesses that rent AI computing power.

Until now, fixing this required either one central boss controlling every job in the data center or expensive new network switches that could reroute traffic automatically. Neither is practical in a shared cloud, where different customers come and go and no single user owns the infrastructure. That's where REACT, a software tool built by researchers, comes in. It quietly attaches to the standard software layer that manages AI communication, watches how long each data exchange takes, and changes the pattern on the fly — like choosing a different teammate to collect the notes when the usual path is jammed.

REACT doesn't change what information is exchanged, just the route it takes. Any individual user can turn it on without asking the cloud provider for help. When tested on a real academic GPU cluster, REACT improved training speed by 13 to 38% even when other users were congesting the network. Simulations under various traffic jams showed gains as high as 75%.

The practical impact is significant: businesses and researchers who rent GPUs pay by the hour. Faster training means lower bills, quicker experiments, and faster improvements to consumer AI products. Because REACT works with common communication libraries like NCCL, it could be adopted widely and fairly soon.