[WIP] btinterp vs PrefixIndex on #21778 base — perf experiment - #21865
Closed
sudeepdino008 wants to merge 29 commits into
Closed
sudeepdino008 wants to merge 29 commits into
sudeepdino008 wants to merge 29 commits into
Conversation
New PrefixIndex type — drop-in replacement for BpsTree lookups.
Per-prefix bucket architecture with adaptive node filling.
- [65536]prefixBucket{firstDI, endDI, nodes} — O(1) prefix lookup
- Adaptive: supplementary scan fills empty buckets with middle key
- narrowWithNodes: per-bucket binary search (max 8 nodes)
- Exact match shortcut: skip disk search on node cache hit
- L1 computed from L2 at build time
Benchmarks (same .kv file, random access):
100K keys: Seek 53% faster, Get 62% faster
1M keys: Seek 46% faster, Get 37% faster
Co-Authored-By: Shuo <shuo@erigon.dev>
PrefixIndex is only built when ERIGON_USE_PREFIX_INDEX=true. Default: false (BpsTree only, current behavior). Usage: ERIGON_USE_PREFIX_INDEX=true ./erigon ... Co-Authored-By: Shuo <shuo@erigon.dev>
Fix F3: Get returns (nil, false) instead of ErrBtIndexLookupBounds for non-existent key in last bucket where endDI==count. Fix F4: addNode copies key bytes via common.Copy to prevent mutation risk from external nodes slice. Add 11 correctness tests: boundary keys, exact match, non-existent, concurrent reads, BpsTree comparison, node key stability. Co-Authored-By: Shuo <shuo@erigon.dev>
…h/erigon into awskii/prefix-index-standalone
Lint fixes:
- Replace bytes.Repeat([]byte{0}, N) with make([]byte, N) (gocritic zeroByteRepeat)
- Replace string comparison with bytes.Equal (gocritic stringXbytes)
- Remove trailing newline (gofmt)
Bench fixes:
- Add missing buildBtreeIndex calls in BenchmarkPrefixIndexGet,
BenchmarkSeekComparison, and BenchmarkGetComparison so the .bt
index file exists before OpenBtreeIndexAndDataFile is called
- Add dbg.UsePrefixIndex=true in BenchmarkPrefixIndexSeek so
bt.search (PrefixIndex) is initialized
Co-authored-by: shuo <shuo@erigon.dev>
…tandalone # Conflicts: # db/datastruct/btindex/bpstree_bench_test.go
…tion - Remove keyCmpFunc callback from PrefixIndex struct and constructors - Add compareKey() method using seg.Reader.MatchCmp directly (zero-copy) - Simplify Seek() binary search: no error return from compare path - Fix comparison direction (cmp < 0 was inverted) - Update tests to match new NewPrefixIndex/NewPrefixIndexWithNodes signatures
Search the leaf window by interpolation instead of binary search: estimate the target from the bound keys' bytes after their common prefix. Probes stay clustered near the target, so cold reads fault far fewer distinct .kv pages. Default on (BT_INTERP=true); falls back to binary after BtInterpBudget (8) probes. No index/format change; M unchanged. Cold btnav us/op (M=256, mainnet 0-8192): storage 171->108, accounts 147->94, code 258->109. The win is page locality, not fewer probes. Bound buffers are stack-backed so the probe loop is allocation-free. TestInterpEquivBinary (fixed keys) and TestInterpEquivBinaryVarLen (variable- length, commitment-style) assert interp == binary for hits+misses across budgets. Same change as erigontech#21794 (-> performance), here targeting main.
Add a footer-based .bt layout: body (nodes + EF) followed by a metadata payload and a fixed 16-byte anchor (footer_len | flags | format_version | magic), with magic last so opens fail fast on a wrong/truncated file. Build the dense Elias-Fano offsets eagerly off-heap from KeyCount+MaxOffset instead of buffering every key through ETL, and recompute each node's data index as di=i*M rather than storing it. A non-zero leading byte distinguishes footer files from the legacy EF-first format.
…adable by old Erigon)
…port Node.off, fix build-sig calls
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
WIP / experiment — not for merge. Integrates three upstream branches to benchmark
.btnavigation methods head-to-head on the new streaming.btformat.Base: #21778 (
alex/bt_streaming_build3_36) — streaming/footer.btformat, M-in-file, off-heap EF build.On top:
BT_INTERP, default on).ERIGON_USE_PREFIX_INDEX).Conflict resolution: kept #21778's footer format; ported PrefixIndex's
Node.offas a runtime-only cache (not encoded); fixed the newBuildBtreeIndexWithDecompressor(existenceFilterPath)signature at stale call sites. Fullbtindextest suite passes (incl.TestInterpEquivBinary, prefix tests).Bench (scratch/btcmp, M64, cold = vmtouch -e, 15k sampled keys, full Get)
SMALL — commitment 9085-9086 (1M, 0.5 GB):
BIG — commitment 8192-8704 (141M, 40 GB), .bt = 235 MB (all three):
Findings
Openis O(.kv size) — a 2x full.kvscan (52 s on the 40 GB file). It warms the whole.kv, so if the entire.kvfits RAM lookups are ~0-fault/7 us (its 37-62% warm win). But with a genuinely cold.kvit is the worst (9.52 faults / 761 us — scatters more than binary), plus the catastrophic open..btsize and heap are comparable across all three.Net: for big/cold commitment under memory pressure (the #21795 regime), btinterp is the clear winner; PrefixIndex only wins when the whole
.kvstays resident.