Optimizing My Disk Usage Program

9 months ago

The article discusses performance improvements in 'dumac', a fast disk usage program for macOS.
The initial version used Tokio for concurrency, but switching to Rayon improved performance by ~28% by reducing thread scheduling overhead.
Rayon's work-stealing design allowed for more efficient recursive directory size calculations with minimal overhead.
The article highlights the reduction in syscalls and context switches (~80% decrease) after switching to Rayon.
A sharded hash-set was used to deduplicate inodes, but inodes' sequential nature caused lock contention.
Optimizing the shard selection by shifting inode bits (inode >> 8) reduced lock collisions from ~176 to ~4 on average.
The final performance gain from reducing inode lock contention was ~5%.
The improvements were benchmarked and verified using hyperfine and macOS Instruments.

Hasty Briefsbeta