Analyzing File Access Characteristics for Deep Learning Workloads on Mobile Devices

Recent advances in artificial intelligence technologies have led to a significant increase in deep learning workloads on mobile devices. Given the limited resources of smartphones, much of the research in mobile deep learning has concentrated on offloading these workloads to edge or cloud servers. While computing resources are crucial, storage I/O remains a critical performance bottleneck for mobile devices, yet the file access characteristics of deep learning have not been thoroughly explored. This paper investigates the file access traces of deep learning workloads on mobile devices, comparing them to traditional workloads. The main findings include: 1) Write access constitutes 48-94% of total file accesses, aligning with conventional mobile apps but contrasting with most desktop applications; 2) Write access in mobile deep learning workloads exhibits repetitive long-loop patterns, offering insights for enhancing file cache performance; 3) Despite its prevalence, write access demonstrates low access skewness; 4) Frequency of file accesses proves more informative than recency in predicting re-access likelihood. The insights from this study are expected to guide the efficient management of future smartphone systems by addressing the unique file access dynamics of deep learning.