superdirstat

WinDirStat alternative for Linux, as a single file that writes a single file.

./superdirstat.py /usr report.html

You get one self-contained HTML report — a squarified treemap with cushion shading next to a synchronised file tree — that you can open with file://, keep, or email. No server, no build step, no dependencies.

It scales to a whole server: 12.1 million files over 3.35 TiB, scanned in 105 seconds into a 1.6 MiB report — about 115,000 files a second, with no index, no daemon and nothing left running afterwards.

Why this exists

ncdu, gdu and dust are excellent, and they live in your terminal. filelight and qdirstat draw the map you want, and they need a desktop session. Neither gives you an artefact: a file you can produce over SSH on a headless box, open a week later, or send to someone who will not install anything.

That is the whole point of this tool. The Python side only walks the tree; the browser does the drawing.

Install

Requires Python 3.7+ (developed and tested on 3.9). Nothing else — the standard library does the scan.

curl -O https://raw.githubusercontent.com/DanXk/superdirstat/main/superdirstat.py
chmod +x superdirstat.py
./superdirstat.py ~/Downloads report.html

Or, for a superdirstat command on your PATH:

pipx install superdirstat

Usage

superdirstat [options] path [output]

Option	What it does
--apparent	Use apparent size (st_size) instead of real disk usage
-x, --one-file-system	Do not cross mount points
-e GLOB, --exclude GLOB	Skip entries matching this glob, tested on both name and full path (repeatable)
--max-nodes N	Maximum nodes in the report (default 80000) — this is what governs the file size
--min-share RATIO	Aggregation threshold as a fraction of the total (default 2e-6, about two pixels)
--min-file BYTES	Aggregate files below this size during the scan, to save memory on huge trees
--include-pseudo	Include /proc, /sys, tmpfs… (skipped by default)
-q, --quiet	No progress output

Scanning / is a supported, ordinary case:

sudo ./superdirstat.py / root.html
# Scanned in 13.1 s: 1,258,492 files, 229,951 folders, 87.4 GiB
# Report: root.html (1.96 MiB, 70,075 nodes)

Reading the report

• Treemap — one rectangle per file, area proportional to size, colour by extension. Cushion shading (van Wijk) is what makes the nesting readable rather than a flat mosaic. Hover for the full path, click to select, double-click a folder to zoom, Esc to go back up.
• File tree — sizes, share of the parent, share of the total. Hovering a row outlines the matching rectangles; hovering a folder outlines its whole region.
• File types panel — every extension with its size and file count. Click one and the rest of the map fades out, leaving only the squares of that type. Esc clears it.
• Right-click menu — on a rectangle or a tree row, copies a shell command for that entry: its path, cd or ls -la on the folder holding it, and for a file, rm -i -- on it. A folder offers a rescan of itself and a find of its largest files instead. Each entry displays the command it copies, and nothing is ever run by the report.
• Theme and language — light or dark, English or French, both remembered. The language selector rewrites the interface only; nothing is re-scanned.

Things worth knowing

Sizes are disk usage, not file length. st_blocks * 512, like WinDirStat, so a sparse or tiny file costs what it really costs. --apparent switches to st_size.

A folder's own blocks are shown. A directory entry occupies disk too, and that space belongs to no child. It appears as a folder overhead leaf, because leaving it out lets children share an area larger than their real sum — worth up to 17 % of distortion on deep, mostly-empty trees.

The tree is pruned, and the report says so. Everything too small to be visible is folded into a + N items leaf; the threshold rises automatically until the node count fits under --max-nodes. The header states the resulting threshold. Without this, a scan of a few million files produces a payload no browser will open.

Virtual filesystems are skipped. /proc, /sys, tmpfs and friends occupy no disk. Their mount points are read from /proc/self/mountinfo rather than hard-coded, because the real list depends on the machine — containers, cgroup v1 vs v2, bind mounts.

Symlinks are never followed; hardlinks are counted once. A symlink's own blocks are counted where the link sits, as du does — what is not done is walking through it. Both counts appear in the header, so a total that differs from another tool is explainable rather than mysterious.

Depth is capped at 2048 levels. No real path goes that deep — PATH_MAX runs out first, at roughly 2000 levels of single-letter directory names — so this only ever fires on a tree something has mangled. When it does, the folders left out are counted in the header instead of taking the scan down with them.

Performance

On a busy server, cold cache excluded:

Tree	Scan	Report	Render in browser
221k files, 9 GiB	2.1 s	1.9 MiB	37 ms
1.26M files, 87 GiB	13.1 s	2.0 MiB	40 ms
12.1M files, 3.35 TiB	105.6 s	1.6 MiB	not measured

That works out at 115,000 files a second, or 136,000 directory entries counting folders, from a single process with no threads: one readdir batch per directory and one stat per entry, which is the least the sizes can be known from. Tools that thread the walk overlap those same calls and can beat it on wall clock; this one buys its speed by not doing anything else. The last row was scanned on a different machine, so its render time is missing; it holds fewer nodes than the row above it, which bounds it.

Memory is the limit to watch, not time. The whole tree is held before pruning, at roughly 300 bytes per entry — 65 MB for 220,000 entries, and close to 4 GB for the 14.4 million of that 3.35 TiB scan. --min-file 65536 aggregates small files during the walk when that does not fit, at the cost of detail on them.

Tests

The report's JavaScript can be checked without a browser:

./superdirstat.py /usr /tmp/r.html
node tests/geometry.js  /tmp/r.html    # exact tiling, proportional areas, hit-testing
node tests/highlight.js /tmp/r.html    # the type filter, pixel by pixel
node tests/bench.js     /tmp/r.html    # layout / paint / highlight timings
node tests/context.js   /tmp/r.html    # the commands the right-click menu copies

See tests/README.md — in particular the reason those tests compile the report with vm.runInThisContext and never eval().

Limitations

• Linux only. st_blocks and /proc/self/mountinfo are the two hard dependencies. macOS would need a different mount enumeration; --apparent works anywhere.
• Read-only by design. No delete, no move — the right-click menu composes commands and puts them on the clipboard, it never executes anything, and only a file is ever offered a delete command. A tool that draws your disk and a tool that empties it should not be the same binary.
• A report is a snapshot. There is no watching, no diffing between two reports yet.

License

MIT — see LICENSE.

The cushion shading comes from Jarke van Wijk and Huub van de Wetering's Cushion Treemaps (1999), by way of WinDirStat and KDirStat. The implementation here is independent.