ZTBP:eBPF-Driven Analysis for Improved Random Read Performance in ZNS Devices via DB Clustering

Despite the introduction of advanced storage technologies, NAND-based devices continue to exhibit significantly higher latencies compared to memory technologies like SRAM/DRAM, hindering storage media performance. With the increasing adoption of Zoned Namespace (ZNS) [5] SSD optimizing their performance for random read workloads has a critical challenge to mitigate the inherent latency limitations of NAND devices. This paper presents ZTBP, a novel framework that leverages the eBPF (Extended Berkeley Packet Filter) [10]-[13] technology to trace and analyze NVMe IO operations on ZNS devices, enabling dynamic cache optimization and improving random read performance. The ZTBP framework employs eBPF for comprehensive NVMe IO tracing and integrates machine learning libraries for applying DB clustering algorithms to identify critical high-read-intensity zones within the ZNS SSD. This approach facilitates prioritized caching of frequently accessed zones, resulting in enhanced cache hit ratios and reduced random read latencies, addressing the latency bottlenecks in NAND-based storage media. Extensive evaluations using expected ZNS workloads and demonstrate the effectiveness of ZTBP in optimizing storage performance and mitigating the latency limitations of NAND devices. The proposed framework achieves substantial improvements in IOPS, bandwidth, and cache hit rates compared to baseline scenarios, with performance gains ranging from 14% to 40 % across various workload patterns and intensities. This research introduces a novel approach to address the challenges of random read performance and latency bottlenecks in NAND-based ZNS SSDs by leveraging eBPF for IO tracing and machine learning for cache optimization. The ZTBP framework contributes to the advancement of storage technologies, enabling more efficient and high-performance storage solutions that overcome the inherent limitations of current NAND devices. This study significantly diverges from previous works by specifically targeting NVMe IO operations analysis for performance enhancement, which wasn't the primary focus of the earlier studies on optimizing RocksDB [7] for ZNS SSDs, the benefits of ZNS interfaces [8], or general performance characteristics of ZNS SSDs. [9]