[jaekwon]

The greatest gas cost in gno.land is disk IO, and the iavl tree in particular. The B+32 tree is a drop-in replacement for the IAVL tree with comparatively short Merkle proofs with drastically less disk reads/writes.

---

Implements tm2/pkg/bptree, a versioned persistent B+ tree with branching factor 32, as a complete alternative to tm2/pkg/iavl for blockchain state storage. Includes a
tm2/pkg/store/bptree CommitStore wrapper that satisfies the same interfaces as the IAVL store.

The core motivation: IAVL's binary tree has ~28 levels for 100M items, meaning 28 disk reads per GET and 28 COW'd node writes per SET. The B+32 tree has ~6 levels for the
same data — ~4x fewer disk operations per mutation.

Key design choices

• B=32 branching factor — 6 levels for 100M items, all inner nodes fit in a 43MB cache
• Out-of-line values — leaves store key + SHA256(value), values stored separately by content hash
• Mini merkle per node — binary merkle over 32 slots with RFC 6962 domain separation (0x00/0x01/0x02), maps cleanly to ICS23 InnerOp chain
• Sentinel short-circuit — empty subtrees at all depths hash to the same sentinel, enabling ICS23 EmptyChild for non-membership proofs
• No leaf sibling pointers — stack-based iteration avoids COW cascade on splits
• 90/10 split for sequential keys (leaves stay ~97% full vs 50% with naive splits)
• Lazy node loading — LoadVersion loads only the root; children loaded on demand via sync.Mutex-guarded getChild
• Dual-tree-walk pruning — no orphan index; compares child hash sets between adjacent versions

Disk I/O comparison (100M items, 10K node cache)

┌───────────────┬──────┬──────┬───────────┐
│ Operation │ B+32 │ IAVL │ IAVL+fast │
├───────────────┼──────┼──────┼───────────┤
│ GET reads │ 3 │ 15 │ 1 │
├───────────────┼──────┼──────┼───────────┤
│ SET reads │ 2 │ 15 │ 15 │
├───────────────┼──────┼──────┼───────────┤
│ SET writes │ 7 │ 28 │ 29 │
├───────────────┼──────┼──────┼───────────┤
│ SET total ops │ 9 │ 43 │ 44 │
└───────────────┴──────┴──────┴───────────┘

Full analysis in PERFORMANCE.md covering cold/warm cache, batched workloads, and bytes-per-SET tradeoffs.

Files

• tm2/pkg/bptree/ — 25 source files, 318 tests
• tm2/pkg/store/bptree/ — CommitStore wrapper with Query, Iterator, proof integration
• tm2/pkg/store/rootmulti/proof.go — registers BptreeCommitmentOpDecoder alongside IAVL

Test plan

• go test ./tm2/pkg/bptree/ ./tm2/pkg/store/bptree/ -race — 318 tests, zero races
• Verify IAVL tests still pass: go test ./tm2/pkg/iavl/
• Verify store tests still pass: go test ./tm2/pkg/store/...
• Review PLAN.md for resolved design questions
• Review PERFORMANCE.md for methodology and assumptions