🤖 AI Summary
This study addresses the significant time and energy overhead (0.25%–46.75%) incurred by existing tools for high-frequency RAPL-based power monitoring, which rely on system calls and frequent polling. Through two controlled experiments evaluating seven tools at a 1 kHz sampling rate, the authors develop lightweight user-space applications and kernel modules that directly access model-specific registers (MSRs) using low-level instructions such as rdmsr, bypassing the high-overhead /proc interface. Their findings reveal that system calls are substantially slower than rdmsr, which in turn is slower than common instructions like cpuid. Based on these insights, the work proposes design principles emphasizing architectural simplification and preferential use of low-level instructions, thereby reducing monitoring overhead to near-baseline levels and enabling efficient high-frequency energy analysis.
📝 Abstract
The Running Average Power Limit (RAPL) interface is widely used to estimate software energy consumption via CPU and DRAM counters, but tool design differences and high-frequency polling can introduce measurement overhead, namely, extra time and energy consumed by the tool itself.This paper quantifies the impact of RAPL-based tools on high-frequency (1 kHz) energy monitoring and investigates mitigation strategies. We conduct two controlled experiments: the first evaluates seven tools, including a user-space application and a kernel module developed by the authors, against a no-tool baseline, using six NAS Benchmark functions to quantify overhead. The second experiment isolates and times key functions for polling Model-Specific Registers (MSRs) (rdmsr and sys/proc_read) to estimate their execution latencies and identify potential slowdowns. The results show that existing user-space tools can introduce substantial time overhead at 1 kHz, whereas our tools significantly reduce system call overhead and inline math overhead. The time overhead of existing tools ranges from 0.25% to 46.75%. Our solutions maintain time overhead levels close to the baseline. We also find that system calls are slower than rdmsr, which in turn is slower than traditionally long-running instructions like cpuid. These findings indicate that RAPL-based energy measurement can be substantially improved by simplifying tool design and employing lower-level instructions to access RAPL values. Our findings provide guidance for practitioners on how to develop high-frequency energy profiling tools, show possible situations that can skew energy values, and demonstrate that access to RAPL values can be faster using specific techniques.