Immersion cooled environmental monitoring and prediction system for the meerKAT imager

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title

University of Cape Town

This paper reports on an immersion cooling method used to reduce power consumption in a data centre. This study involves a case study of the MeerKAT Science Processor that is responsible for the MeerKAT imaging pipeline. Immersion cooling brings a coolant into direct physical contact with the chips and the circuit board by directly immersing computing equipment into a bath of cooling fluid. According to the National Security Agency's Laboratory for Physical Sciences (LPS), who acquired and installed an oil-immersion cooling system in 2012, the use of immersion cooling means that much of the infrastructure needed for cooling a data centre can be eliminated; it can moreover reduce server failures, and is cleaner and quieter than air cooling. In this study, an oil cooled environment is created and a prototype low-cost thermal management system for the system is built and tested. This prototyped system was called the Environmental Monitoring System (EMS), and it monitors humidity and temperature of the oil-cooled environment. In this study, discrete temperature data gathered by the thermal management system is used to build a prediction program that we called the Immersion Cooling Temperature Predictor (ICTP), which predicts temperature at locations not covered by sensors. The ICTP predicted measurements using a Gaussian process model, providing estimates for non-sampled locations to help make the monitoring systems more fault tolerant. Reducing the number of sensor nodes moreover reduces installation costs, as well as space utilized and power consumed by temperature sensors. The accuracy of detecting hotspots in an immersion cooled environment using the system is also investigated. From the experiments it was found that the ICTP had a mean error of 0, 0083, standard deviation of 2, 56 and predicted standard deviation of 2, 44 for predicting hotspots.