perf(dash-spv): reduce ram usage

[ZocoLini]

Reduce dash-spv RAM usage

Draft: the measurements below were taken with #1015, #1016 and #1014 applied (the committed-range sweep removal and its prerequisites). This branch does not depend on them to build or pass tests, but it stays a draft until they merge and until the on-disk format change below is handled.

Peak RAM, mainnet restore

dash-spv-bench, scenario mainnet.100mbi.100ms, one wallet. Every run ends with the same wallet state: 14114383 sat, 7112 transaction records, 13389 addresses.

Build	Peak RSS	Sync time
dev (committed-range sweep still in)	2280–2360 MiB (jemalloc), ~2924 MiB (glibc)	20–24 min
dev + #1015 + #1016 + #1014	982 MiB (jemalloc)	8.2 min
dev + #1015 + #1016 + #1014 + this branch	538 MiB (jemalloc), 547 MiB (glibc)	7.5–8.0 min

With the sync fixes applied, this branch roughly halves the peak with no cost in sync time.

jemalloc numbers come from --memory-snapshot runs, which link jemalloc with heap profiling. glibc numbers come from plain runs with the default allocator.

Changes, in commit order

• Bench: in-process profiling, client always in docker.

• run.sh --flame builds with the cpu-profile feature (pprof) and writes flamegraph.svg.

• --memory-snapshot builds with the heap-profile feature (jemalloc) and writes heap-peak.svg, the live heap at the highest RSS of the run.

• Neither profiler is compiled in without its flag.

• The client always runs in its container, built inside the Rust image.

• Download shaping requires ifb; the policer fallback is gone.

• run.sh goes from 664 to 366 lines.

• Drop the committed-height segment release (reverts fix(dash-spv): release clean storage segments below the committed height during long scans #946). Memory was only freed when a sync manager called set_committed_height, and only on the next timed persist. The segment cache now bounds its resident set by itself.

• Trace segment cache hits and misses. One trace-level line per segment request says whether it was a hit, or where a miss came from (disk or new). Enable it with RUST_LOG=dash_spv::storage::segments=trace.

• Keep at most two segments resident per storage cache (was ten). Each cache still serves ~2.34M requests from memory per restore.

• Persist a dirty segment when it is evicted. Dirty segments no longer wait in an in-memory evicted map for the 5 s storage tick, so the resident limit is a real limit.

• Stream segments to disk. Segment::persist used to encode the whole segment into one Vec, up to 128 MiB for a block segment near the tip. It now writes item by item through a BufWriter into the temporary file, which is then synced and renamed as before.

• Size storage segments per item type (Persistable::ITEMS_PER_SEGMENT, previously 50 000 for every type):

  Type	Items per segment	Segment size
  Headers	10 000	~1.1 MB
  Filter headers	50 000	~1.6 MB
  Filters	2 000	~2 MB near the tip
  Blocks	1 000	—

  • Segment caches drop from 332 MiB to 38 MiB at the peak.
  • With 1 000 items for every type, the header and filter header phases paid an fsync per evicted segment. Sizing per type avoids that.
  • Keep Persistable inside the storage module.

On-disk format

The per-type segment sizes change the layout of the header, filter and block segment files. With this branch, storage written with 50 000-item segments is misread. Before this leaves draft it needs either a versioned folder (forcing a re-sync) or a migration on open.

Testing

• cargo clippy --all-features --all-targets -D warnings.
• dash-spv lib tests, the dash-spv dashd integration suite, and dash-spv-ffi dashd_sync.
• Mainnet bench restores as in the table above, with an identical wallet in every run.

What remains in the heap

From the jemalloc heap profile at the peak:

• Header hash→height index: ~160 MiB.
• Wallet: ~60–90 MiB.
• Masternode lists: ~65–80 MiB.
• Segment caches: ~40 MiB.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

❌ Patch coverage is 96.36364% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 77.17%. Comparing base (350f5f7) to head (3a87c3a).

Files with missing lines	Patch %	Lines
dash-spv/src/storage/segments.rs	89.47%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##              dev    #1024      +/-   ##
==========================================
- Coverage   77.23%   77.17%   -0.07%     
==========================================
  Files         329      329              
  Lines       83862    83686     -176     
==========================================
- Hits        64768    64581     -187     
- Misses      19094    19105      +11     

Flag	Coverage Δ
core	78.24% <ø> (ø)
ffi	50.94% <ø> (ø)
rpc	20.00% <ø> (ø)
spv	92.05% <96.36%> (-0.13%)	⬇️
wallet	79.79% <ø> (ø)

Files with missing lines	Coverage Δ
dash-spv/src/storage/blocks.rs	100.00% <ø> (+1.16%)	⬆️
dash-spv/src/storage/filters.rs	100.00% <ø> (+1.61%)	⬆️
dash-spv/src/storage/io.rs	97.68% <100.00%> (+0.46%)	⬆️
dash-spv/src/storage/mod.rs	86.15% <ø> (+1.04%)	⬆️
dash-spv/src/sync/blocks/manager.rs	96.72% <ø> (-0.07%)	⬇️
dash-spv/src/sync/filters/manager.rs	97.80% <ø> (-0.01%)	⬇️
dash-spv/src/storage/segments.rs	96.52% <89.47%> (-0.21%)	⬇️

... and 4 files with indirect coverage changes