Facilitating the HPC Data Center Host efficiency through Big Data Analytics

Quality of service is important feature for a High Performance Computing Center (HPC) center like Partnership for an Advanced Computing Environment (PACE) center in Georgia Institute of Technology (Georgia Tech). The user's job fails running on a HPC center may due to a spectral of reasons, one of major contributor is the hardware and network failure. Reducing the hardware failure rate can significantly increase a data center's quality of service as well as reducing the cost of human intervention. This is critical during PACE's transition to a fee-based service model in which uptime correlates directly with revenue. PACE has around 9 millions jobs each year with 12% of job failure rate. In order to extend service life of hardware and reduce the potential failure and data center's cost, we present a machine learning method to understand the center's host usage pattern. By clustering the hosts based on multiple features, we reshuffle the host list to avoid the hosts being overused over time. We build a test framework which runs the complex combination of experiments, and presents the ad hoc comparisons. We intend to make the machine learning method in a rack aware fashion, and show the meaningful result with rack information included.