Dr. Brian Setz
- e-mail: briansetz [at] gmail.com
- personal website: Brian Setz
Research
- Service-Oriented Architecture
- Wireless Sensor Networks
- Data Center Efficiency
Recent publications
- Enough Hot Air: The Role of Immersion Cooling ( ), In Energy Informatics, 2023.
Abstract
Air cooling is the traditional solution to chill servers in data centers. However, the continuous increase in global data center energy consumption combined with the increase of the racks’ power dissipation calls for the use of more efficient alternatives. Immersion cooling is one such alternative. In this paper, we quantitatively examine and compare air cooling and immersion cooling solutions. The examined characteristics include power usage efficiency (PUE), computing and power density, cost, and maintenance overheads. A direct comparison shows a reduction of about 50% in energy consumption and a reduction of about two-thirds of the occupied space, by using immersion cooling. In addition, the higher heat capacity of used liquids in immersion cooling compared to air allows for much higher rack power densities. Moreover, immersion cooling requires less capital and operational expenditures. However, challenging maintenance procedures together with the increased number of IT failures are the main downsides. By selecting immersion cooling, cloud providers must trade-off the decrease in energy and cost and the increase in power density with its higher maintenance and reliability concerns. Finally, we argue that retrofitting an air-cooled data center with immersion cooling will result in high costs and is generally not recommended.
BibTeX
url - CO2 Emission Aware Scheduling for Deep Neural Network Training Workloads ( ), In 2022 IEEE International Conference on Big Data (Big Data), IEEE, 2022.
Abstract
Machine Learning (ML) training is a growing workload in high-performance computing clusters and data centers; furthermore, it is computationally intensive and requires substantial amounts of energy with associated emissions. To the best of our knowledge, previous works in the area of load management have never focused on decreasing the carbon emission of ML training workloads. In this paper, we explore the potential emission reduction achievable by leveraging the iterative nature of the training process as well as the variability of CO 2 signal intensity as coming from the power grid. To this end, we introduce two emission-aware mechanisms to shift the training jobs in time and migrate them between geographical locations. We present experimental results on power and carbon emission of the training process together with delay overheads associated with emission reduction mechanisms, for various, representative, deep neural network models. The results show that following emission signals, one can effectively reduce emissions by an amount that varies from 13% to 57% of the baseline cases. Moreover, the experimental results show that the total delay overhead for applying emission-aware mechanisms multiple times is negligible compared to the jobs’ completion time.
BibTeX
urldoi - Mining Sequential Patterns for Appliance Usage Prediction ( ), In International Conference on Smart Cities and Green ICT Systems, 2018.
BibTeX
- Planning meets activity recognition: Service coordination for intelligent buildings ( ), In Pervasive and Mobile Computing, Elsevier, volume 38, 2017.
- Metrics for Sustainable Data Centers ( ), In IEEE Transactions on Sustainable Computing, 2017.
BibTeX
(For more publications go to Brian's publication page)