Comparing the Performance of Numba and CUDA for Historical Analysis of Synchrophasor Data

Modern Wide Area Monitoring systems (WAMS) incorporating Phasor Measurement Unit (PMU) technology are producing big datasets. Historical analysis of PMU data is beneficial in development of online WAMS applications, quantifying baseline normal performance, and discovering anomalous events. Energy and time-efficient computational techniques are beneficial for historical analysis of PMU data. Application workflows that include historical analysis typically combine higher-level (but slow) languages like Python with faster (but older) languages like C. This paper compares the performance of Numba Python and C for historical analysis of PMU data, on both the CPU and GPU. We augment a known PMU anomaly detection scheme with linear state estimation, implement it separately in Numba and C, test the approaches on two real-world datasets, and measure their performance on the CPU and GPU of the NVIDIA Jetson Xavier single board computer, varying the available power modes. Results demonstrate that while Numba is significantly faster than traditional Python, simplifies application development, and holds promise for PMU applications, there is a noticeable performance gap between Numba and C on the GPU.