Immersion Cooling Monitoring and Prediction System for Meerkat Radio Astronomy Imager

This paper reports on the development of an immersion cooling monitoring and prediction system to assist in the detection of hotspots (i.e., potential points of failure due to overheating) in an immersion cooled computing cluster, and to assist in scheduling maintenance of these resources in a data centre. Our work focuses on a case study involving resources used in the Science Processor of the MeerKAT radio astronomy array instrument. This system relates to the Square Kilometre Array (SKA) data centre situated in the Karoo, a semi-desert region in South Africa that experiences high daytime temperatures for which traditional air-cooling solutions are inefficient. In this study, an oil-cooled environment and thermal management system was prototype and tested. This prototyped system was called the Environmental Monitoring System (EMS), and monitors humidity and temperature of the oil-cooled environment. Temperature data gathered by the EMS was used to build a prediction program, namely the Immersion Cooling Temperature Predictor (ICTP). The ICTP predicts temperatures between sensors, to reduce the number of sensor nodes and consequently installation and power costs. We report on the accuracy of detecting and predicting hotspots in the environment. From our experiments the ICTP had a mean error of 0.0083oC between predicted and actual measurements within 5 cm from a physical sensor (for the largest vat tested), with standard deviation of 2.44°C. The predictions become increasingly less accurate for points further from a sensor.