v1.0.1 — Disable seq_bypass, 18% SQLite full scan gain

Disables the seq_bypass optimization which harmed real-world workloads.

Benchmarks (4GB SQLite, full table scan, cold cache, interleaved OFF/ON):

• OFF avg 52.58s → ON avg 43.12s → 18.0% speedup
• 3 passes, module reloaded between each OFF/ON pair
• hit_rate_real: 99%, stride_predictions: 5.74M per pass

Changes:

• Commented out seq_bypass return statements in nsd.c (stride-bypass block, workload update, hot update)
• Code comments translated to English
• SQLite full-table-scan benchmark script added

NSD v1.0.0 - Initial Public Release

NSD v1.0.0 - Initial Public Release

Neural Storage Driver — Linux kernel module that learns I/O access patterns at runtime (Markov, Bigram, Stride, FreqRecency) and prefetches data using vfs_fadvise(WILLNEED).

Performance

Device	Workload	Gain
SSD (NVMe)	Sequential 64K	+20.5% (390 → 470 MB/s)
HDD	Sequential 64K	−4% → 0% (regression fixed)
HDD	Random 4K	~0%

Architecture

• VFS-level kprobe (vfs_read) + vfs_fadvise — FUSE/ntfs-3g compatible
• Per-CPU lock-free ring buffer + worker thread
• Device-class auto-tuning (NVMe/SSD/HDD)
• Stride bypass, death-spiral recovery, hysteresis

License

Dual-license: GPL-2.0-only OR Commercial
Commercial licensing: nsd.project.dev@gmail.com

Installation

make
sudo insmod nsd.ko
cat /sys/kernel/nsd/stats

Testing

sudo ./scripts/run_tests.sh [ssd|hdd|hdd_deep|all]