Implementation of green computing in IBM HPC software stack on accelerator based super computing
Nagarajan Kathiresan, Ganesan Narayanasamy · IEEE International Conference on High Performance Computing, Data, and Analytics · 2012
Computational requirements of the high performance computing servers and their power consumptions have increased over the period of time. The major power consumption components are processors and the main memory in the servers. The multi-core era bring the quad, hexa or an octa cores into a single socket, the total power consumption of the server is reduced by the way of various energy aware kernel implementations. I.e., the energy aware kernel will bring the processor/core into sleep modes (or) lower clock speeds during the time of CPUs are not utilized for a finite interval of time. Additionally, the green and energy-aware implementations are done by the way of (i) processor affinity or binding tools (ii) energy aware scheduling techniques (iii) case study by power and performance trade-offs (iv) On-demand compute awareness in the kernel (v) dynamic frequency scaling etc. into the software stack. Herewith, we discussed the various possibilities of energy-aware implementations in the software stack thereby the best application performance and the optimal power/energy efficiency can be obtained within a given power budget and using various user defined performance metrics & constrains.