DISK USAGE, WITHOUT THE WAIT.
Find the folders filling your disk and review cleanup candidates. Native metadata reads and parallel traversal share one Rust engine, available from the CLI, GUI, and AI agents.
v0.5.0-beta.7 · Published beta · MIT
$ hyperdu . --top 3
Top 3 under . (physical desc):
1. . | phys=18.2 GiB | log=17.9 GiB | files=412088
2. ./target | phys=12.4 GiB | log=12.1 GiB | files=48211
3. ./.git | phys=3.1 GiB | log=2.6 GiB | files=1530CLI · GUI · MCP
Two progress lines redraw at speed while it scans, then the largest folders remain.
01 / BENCHMARKS
Linux was compared with GNU du in eight runs per tool on GitHub Actions. Windows measurements add a 1M-file HyperDU-only result and a 10K-file diagnostic. These are records of the displayed benchmark commit, not fresh measurements of the published release or current HEAD. Windows adds a comparison with dua-cli and reference timings for tokei, which does a different job.
AMD EPYC 9V74 80-Core Processor · 4 vCPU · 15.6 GiB RAM · ext4
Ubuntu 24.04 · 1M files per tree · 256 B per file · median of 8 runs per tool · allocated size
| Dataset | Files | HyperDU | GNU du | Speed ratio |
|---|---|---|---|---|
| Flat directoryflat | 1,000,000 | 2327.49 ms | 2763.52 ms | 1.187× |
| Wide directory treewide | 1,000,000 | 745.96 ms | 2428.8 ms | 3.256× |
| Deep directory treedeep | 1,000,000 | 764.67 ms | 2434.28 ms | 3.183× |
AMD Ryzen 9 3900X · 12 cores / 24 threads · 128 GiB RAM · NTFS / NVMe
Windows 11 · 1M files · 256 B per file · logical size · median of 8 HyperDU-only runs
| Dataset | Files | HyperDU | GNU du (MSYS) | Speed ratio |
|---|---|---|---|---|
| Flat directoryflat | 1,000,000 | 589.91 ms | Timed out at 600 s | No comparison |
GNU du timed out after 600 seconds during warmup, before any measured run. The 1M-file comparison is incomplete; wide and deep 1M-file workloads were not run.
AMD Ryzen 9 3900X · 12 cores / 24 threads · 128 GiB RAM · NTFS / NVMe
Windows 11 · 10K files per tree · 256 B per file · logical size · median of 2 runs per tool
| Dataset | Files | HyperDU | GNU du (MSYS) | Speed ratio |
|---|---|---|---|---|
| Flat directoryflat | 10,000 | 32.98 ms | 704.17 ms | 21.35× |
| Wide directory treewide | 10,000 | 27.46 ms | 712.33 ms | 25.94× |
| Deep directory treedeep | 10,000 | 32.32 ms | 836.73 ms | 25.89× |
All directory totals matched between tools. These ratios apply only to this 10K-file, two-run diagnostic, not to 1M-file workloads or the Linux comparison.
AMD Ryzen 9 3900X · 12 cores / 24 threads · 128 GiB RAM · NTFS / NVMe
Windows 11 · logical size · median of 12 runs per tool · three 100K-file synthetic trees + two real trees
| Dataset | Files | HyperDU | dua-cli 2.45.0 | Speed ratio |
|---|---|---|---|---|
| Wide directory tree (synthetic)wide | 99,856 | 67.57 ms | 187.37 ms | 2.77× |
| Deep directory tree (synthetic)deep | 100,000 | 179.68 ms | 194.62 ms | 1.08× |
| Flat directory (synthetic)flat | 100,000 | 98.8 ms | 997.51 ms | 10.1× |
| Cargo registry (real files)cargo-registry | 107,953 | 478.55 ms | 554.61 ms | 1.16× |
| node_modules (real files)node-modules | 27,940 | 72.26 ms | 156.6 ms | 2.17× |
Byte totals from both tools equalled the independent walk on every dataset. HyperDU was faster on all five, but the gap depends on tree shape (1.08-2.77x excluding flat). The deep tree has almost no width to parallelize, so the two are close. The flat figure includes dua printing one row per direct entry (100,001 rows) and is not a representative ratio. Unrelated jobs held the CPU at roughly 60-90% busy during measurement, so absolute times are inflated.
AMD Ryzen 9 3900X · 12 cores / 24 threads · 128 GiB RAM · NTFS / NVMe
Windows 11 · two real trees · median of 12 runs per tool · HyperDU sums sizes, tokei counts lines
| Dataset | Files | HyperDU | tokei 15.0.0 | Speed ratio |
|---|---|---|---|---|
| Cargo registry (real files)cargo-registry | 107,953 | 478.55 ms | 9407.19 ms | No comparison |
| node_modules (real files)node-modules | 27,940 | 72.26 ms | 782.31 ms | No comparison |
tokei reads file contents and counts lines, while HyperDU only totals sizes. These are different jobs, so no speed ratio is shown; the times are reference only. tokei parsed 80,897 files in the Cargo registry and 16,576 in node_modules (recognized source extensions only). The synthetic fixtures have no extensions, so tokei parses nothing there and was not timed.
All values are warm-cache medians. Speed ratio is the comparison tool's time (GNU du or dua-cli) divided by HyperDU time; tokei does a different job, so no ratio is shown. Linux uses allocated size; Windows uses logical size. OS, accounting, file counts and run counts differ, so ratios cannot be compared across environments or extended to unmeasured conditions.
See every run and the method ↗Linux GitHub workflow run ↗02 / UNDER THE HOOD
Change how metadata is read and how work is shared. Platform-specific reads reduce extra operations; per-worker queues and work stealing distribute traversal across uneven directory trees.
NtQueryDirectoryFileBatches names, sizes, allocation sizes, and file IDs to reduce additional file opens while scanning large folders.
NTFS · native CIgetdents64 + statxReads directory entries and required metadata, then aggregates in parallel. It does not eliminate the I/O needed for accurate sizes.
CLI · GNU du comparisongetattrlistbulkBulk metadata reading is implemented, but macOS is outside current CI and release targets. CLI verification remains incomplete.
Implementation · not in current CIOn eligible Windows NTFS volume roots, --mft can try direct reads with the required privileges and options. Incomplete required reads or parsing fall back to enumeration; success does not guarantee complete accounting parity.
Read the architecture ↗03 / YOUR WORKFLOW
Inspect in a terminal, explore on your desktop, or integrate with an app or agent. Choose the interface and reuse the scanning engine.
Rank directories and export JSON or CSV. GNU du compatibility mode is available, but not a complete replacement for every du option.
hyperdu . --top 20
hyperdu . --json usage.json
hyperdu --compat gnu -k /var/logSee completed child folders while the remaining scan runs. The Windows and Linux GUI offers scan modes, filters, and detailed settings.
hyperdu-gui ↗Embed the independent hyperdu-core library in Rust without launching the CLI. Reuse scanning, aggregation, and platform optimizations.
hyperdu-core ↗04 / FOR YOUR AGENTS
From free space to large folders and cleanup candidates, hyperdu mcp provides typed tools and structured results.
All three MCP tools are read-only and never delete files. A candidate is not proof that removal is safe; a person reviews and decides. The Skill also works through the CLI alone.
Set up MCP / Skill / Plugin ↗list_volumesWhich volume is running out of space?
scan_pathWhat is taking up that space?
find_reclaimableWhich outputs might be regenerated?
cargo install hyperdu --locked --version 0.5.0-beta.7
codex mcp add hyperdu -- hyperdu mcp05 / GET STARTED
Install the published beta from crates.io. One hyperdu command includes the CLI and MCP server. To run without Rust, choose a Windows or Linux binary from Releases.
cargo install hyperdu --locked --version 0.5.0-beta.7Cargo source builds need a recent stable Rust (the CLI declares a 1.88 minimum) and platform build tools. Prebuilt binaries do not require Rust to run.