A treemap for finding and removing what fills your disk, for Omarchy. Rust + GPUI.
SSH indexing profile
Measured on 2026-10-22, Intel Core Ultra X7 358H (16 CPUs), release builds. All fixtures were disposable local tmpfs trees with warm metadata. These are component measurements, not a profile of a6000's storage or UI. Decode timings include tree destruction at the end of each invocation. A read-only process check found no active helper there; no new real-server scan was started. No caches were dropped and no limits were raised.
Results
| Component / fixture | Before | After |
|---|---|---|
| Decode + aggregate + classify 500,000 wire records | 288.3 ± 13.7 ms | 210.2 ± 2.2 ms |
| Stream 100,000 files, output discarded | 103.4 ± 2.7 ms | 101.2 ± 2.0 ms |
| Parallel local scan, 100,000 files, earlier paired run | 21.5 ± 0.5 ms | unchanged |
| Serial streaming, same earlier paired run | 86.8 ± 2.6 ms | — |
| Actual static helper + localhost SSH + decode, 100,000 files | — | 321–347 ms uncompressed; 316–334 ms compressed |
| Helper RSS, 70,000 external hardlinks, 20k / 40k records | scan aborted at 65,536 identities | 2,888 / 2,888 KiB; completes |
Hyperfine used warmups and 5–7 trials. SSH used five trials per mode and an isolated sshd, temporary keys/config, the shipped static musl helper and no user SSH settings. SSH measurements include connection/authentication but not automatic deployment. Do not subtract these runs to claim a precise network cost: the implementations differ and work overlaps.
Findings
- Avoidable decoder work.
perf recordshowed genericserde_json::Valueconstruction followed by a second deserialization. Every record also allocated a new line buffer. Direct tuple/control decoding and one reused line buffer reduced this fixture's decode time by 27% (1.37× throughput). This is not a 27% end-to-end scan improvement. - Serial metadata remains the main walker limitation. In a paired fixture, the existing parallel local scanner was 4.04× faster than the serial streaming helper. Streaming time did not materially change after the decoder fix. Bounded metadata concurrency is the next substantial scanner optimization; reverting to the full tree-building local engine on the server would bring back file-count-scaled memory.
- Compression is not the explanation on loopback. The compressed and uncompressed ranges overlap. This says nothing about a slower real link.
- The hardlink failure was our bookkeeping policy, not memory pressure.
nlinkincludes names outside the scanned root. Tracking “remaining names” cannot drain for such files, so ordinary uv/package caches hit the 65,536-entry cap. Modern sessions now negotiateclient_dedup: the helper emits(device, inode)with linked file records and keeps no identity table. The desktop uses the local scanner's existingSeenbitmap/sharded set. In-root aliases are charged once; external names need not be visited. Legacy peers retain the old bounded wire mode.
The after-profile shows tree sorting/classification, typed JSON parsing, line reading and clock checks as remaining desktop costs. Its wide, single-directory fixture exaggerates sorting relative to ordinary trees. A future profile of an already-running real scan should distinguish helper CPU, filesystem wait, SSH throughput and desktop finalization before attributing its wall time to any one of these synthetic results.
Reproduce safely
cargo build -p disktree-core --example profile_remote --release \
--config profile.release.debug=1
target/release/examples/profile_remote generate /tmp/disktree-wire.jsonl
hyperfine --warmup 1 --runs 7 \
'target/release/examples/profile_remote decode /tmp/disktree-wire.jsonl'
perf record -g --call-graph dwarf -- \
target/release/examples/profile_remote decode /tmp/disktree-wire.jsonl
perf report
For an explicitly created disposable fixture under
/tmp/disktree-profile-*, compare scan PATH and stream PATH >/dev/null
with that example. It refuses non-temporary paths; it has no default root.
The SSH profiler uses the existing isolated test harness:
DISKTREE_PROFILE_ROOT=/tmp/disktree-profile-fixture \
bun scripts/test-ssh.ts
That directory must already exist and contain only your disposable test
files. The harness automatically builds the portable helper and requires
the profile_remote example built above. Without the environment variable,
it runs the normal four SSH integration tests instead.
Memory/failure regressions against the actual helper executable:
cargo test -p disktree-scan --test stdio \
external_hardlinks_have_bounded_resident_memory -- --ignored --nocapture
cargo test -p disktree-scan --test stdio \
wide_stream_has_bounded_resident_memory -- --ignored --nocapture
Both create and clean their own temporary fixtures. The regular core suite also verifies more than 65,536 identities, duplicate aliases, external names, disabled de-duplication, legacy records and complete framing.