Job class aware liquid cooling improves ai data center heat control
Job Class Thermal Intent Aware Liquid Cooling Allocation for AI Data Centers
Distributed, Parallel, and Cluster ComputingPerformance
Summary
AI data centers use GPUs that get very hot, and traditional air cooling can't keep up with the heat they produce. The authors created a system called JCTI that uses knowledge about upcoming jobs to get the liquid cooling ready before the GPU temperature rises. This approach reduces overheating events by more than half and cuts excess cooling significantly compared to standard methods. Their system connects scheduling and cooling controls to work together rather than separately.
What this means in practice
- •For data center operators: Adjust cooling flow proactively based on scheduled GPU jobs to reduce overheating and improve efficiency in AI workloads.
- •For energy grid managers: Use improved thermal-aware scheduling data to better predict and smooth power demand from large AI data centers.
Authors
Krishna Chaitanya Sunkara
Abstract
GPU-dense AI data centers need to run on liquid cooling as air simply cannot shed the heat at these power densities. Yet the cooling loops themselves are blind to what workloads are about to run; they crank up flow only after a sensor catches a temperature climb, which can take 30 to 50 seconds. We built Job-Class Thermal Intent (JCTI) to close that window. The scheduler already knows a job is coming and what class it belongs to; JCTI feeds that information straight to the cooling controller so it can stage coolant before the heat shows up. We pulled the thermal signatures for each job class out of MLPerf GPU power traces and tuned arrival patterns against Alibaba cluster data. Over 120 paired Monte Carlo trials the numbers come out to 56.4% fewer thermal violations and 60.2% less cumulative overshoot than a straight PI loop. As AI data centers evolving towards gigawatt grid loads with highly fluctuating power swings, thermally-aware scheduling reduces sudden demand and improves load prediction in grid side. Cooling and scheduling have been running as two separate systems for years despite each one knowing something the other needs, JCTI wires them together.