Dwarf in a Giant: Enabling Scalable, High-Resolution HPC Energy Monitoring for Real-Time Profiling and Analytics
share
Energy efficiency, predictive maintenance and security are today key challenges in High Performance Computing (HPC). In order to be addressed, accurate monitoring of the power and performance, along with real-time analysis, are required. However, modern HPC systems still have limited power introspection capabilities, lacking fine-grain and accurate measurements, as well as dedicated systems for live edge analysis. With the goal of bridging this gap, we developed DiG (Dwarf in a Giant), an enabler framework for green computing, predictive maintenance and security of supercomputers. DiG provides high quality monitoring of power and energy consumption of HPC nodes. It is completely out-of-band and can be deployed in any hardware architecture/large-scale datacenter at a low cost. It supports (i) fine-grained power monitoring up to 20us (50x improvement in resolution than state-of-the-art - SoA); (ii) below 1 power measurements, which makes it suitable for the most rigorous requirements of HPC ranking lists (i.e. Top500); (iii) high-precision time-stamping (sub-microsecond), which is three order of magnitude better than SoA; (vi) real-time profiling, useful for debugging energy aware applications; (v) possibility for edge analytics machine learning algorithms, with no impact on the HPC computing resources. Our experimental results show it can capture key spectral features of real computing applications and network intrusion attacks, opening new opportunities for learning algorithms on power management, maintenance and security of supercomputers.
READ FULL TEXT
