Rear-door heat exchanger with workload-aware coolant valve control for AI server racks
G06F 1/20 Cooling means for computers
G06F 9/50 Allocation of resources
A liquid-cooled rear-door heat exchanger on a server rack has a motorised coolant valve that is opened in advance, based on the job scheduler's queue, rather than only after the rack outlet air has already become hot. When the scheduler is about to start a GPU-heavy job on servers in a rack, it sends the expected power draw to a valve controller, which raises coolant flow 30 to 120 seconds before the load starts. This reduces temperature overshoot at the start of training jobs and lets the coolant supply temperature be set higher on average.
- Rear-door heat exchangers for server racks (general technique, widely sold)
- Liquid-cooling control based on outlet temperature (common practice)
Problem
AI training jobs switch GPU racks from idle to full power within seconds; a rack can go from about 5 kW to over 40 kW. Rear-door heat exchangers are normally controlled by a temperature sensor at the outlet, so the valve only reacts after hot air has already left the rack. The delay causes hot spots, fan speed spikes and thermal throttling, and operators compensate by keeping the coolant colder than needed all the time, which wastes chiller energy.
Detailed description
The system has a rear-door heat exchanger (10) fitted to a server rack (12), a motorised two-way coolant valve (14) on the coolant return line, a coolant distribution unit (16) and a valve controller (18). The controller is a small computer connected to the cluster's job scheduler over the network.
When a job is queued, the scheduler already knows which servers it will use and their typical power draw for that job type. Before starting the job, the scheduler sends the controller a message with the rack ID, the expected start time and the expected power per server. The controller converts the expected rack power into a target coolant flow using a stored curve for that heat exchanger (flow needed per kW at the current supply temperature) and starts opening the valve a lead time ahead of the job start. The lead time is chosen from the measured response time of the heat exchanger, typically 30 to 120 seconds.
After the job has started, the normal outlet temperature loop takes over and corrects the flow. If the job does not start as announced (cancelled or delayed), the controller returns to the temperature-based setting after a timeout, for example 5 minutes. The controller also sends the scheduler the current spare cooling capacity per rack, so the scheduler can avoid placing two large jobs in the same rack at once.
Example
A rack with 8 GPU servers idles at 4.8 kW. The scheduler announces a training job using all 8 servers at about 4.2 kW each (33.6 kW), starting in 90 seconds. With coolant supplied at 24 °C, the stored curve gives about 1.5 l/min per kW, so the controller sets the valve for about 50 l/min, starting 60 seconds before the job. In a test with a temperature-only controller, the outlet air peaked at 41 °C for about 2 minutes; with the advance signal it stayed below 33 °C.
Variants
Instead of a rear-door heat exchanger, the same advance signal can control direct-to-chip cold plate loops, in-row coolers, or the fan speed of air-cooled racks.
The expected power can come from the job's requested GPU count, from past runs of the same job, or from a power model; it can also be sent per server rather than per rack.
The valve can be replaced by a variable-speed pump per rack. Several racks on one coolant loop can be pre-loaded in sequence, so the distribution unit is not asked for all the extra flow at the same moment.
Drawings
Figure 1 shows the rack seen from the side. The heat exchanger (10) is on the rear door of the rack (12). The return line runs through the valve (14) to the coolant distribution unit (16). The controller (18) receives the job scheduler's message and drives the valve (dashed lines).