Abstract
The high energy consumption of HPC systems is an obstacle for ever-growing systems. Unfortunately, energy consumption does not decrease linearly with reduced workload; therefore, energy conservation techniques have been deployed on various levels which steer the overall system. While the overall saving of energy is useful, the price of energy is not necessarily proportional to the consumption. Particularly with renewable energies, there are occasions in which the price is significantly lower. The potential of saving energy costs when using smart contracts with energy providers is lacking research. In this paper, we conduct an analysis of the potential savings when applying cost-aware schedulers to data center workloads while considering power contracts that allow for dynamic (hourly) pricing.The contributions of this paper are twofold: (1) the theoretic assessment of cost savings; (2) the development of a simulator to replay batch scheduler traces which supports flexible energy cost models and various cost-aware scheduling algorithms. This allows to approximate the energy costs savings of data centers for various scenarios including off-peak and hourly budgeted energy prices as provided by the energy spot market. An evaluation is conducted with four annual job traces from the German Climate Computing Center (DKRZ) and Leibniz Supercomputing Centre (LRZ).The theoretic analysis indicates a cost savings for 4–8% when shutting down unused client nodes, and 6–20% with hourly cost models and optimal scheduling. The experimental validation of a practicable scheduler increases the accuracy against the theoretical best case analysis. As expected, a cost-efficient scheduling algorithm that is fed with the information about future energy costs shifts the jobs to the timeslots where the job execution is cheaper and reduces the energy expenditure, yet increases the waiting times of pending jobs. However, the expected savings for this effort are not justifiable compared to the simple strategy of turning off the unused nodes. Additionally, we compare the cost savings to the total costs of ownership showing that smaller systems with on-demand provisioning yield better cost efficiency.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have