For an InternalNode at a group-boundary depth, hashInternal previously
computed pure SHA256(left, right) recursion over the natural-depth
in-memory tree built by UpdateStem. But serializeSubtree extends stems
to the group's bottom layer via key-bit extension, so the on-disk blob
encodes an extended-depth structure. When a fresh reader deserializes
that blob, hashInternal walks the extended-depth in-memory tree and
produces a different value.
The result was that for any subtree with multiple stems sharing a
prefix shorter than groupDepth, the parent's stored child-hash (computed
from the natural-depth in-memory tree at commit time) did not equal the
child blob's read-back hash. Geth's own write-read cycle was internally
inconsistent: state-actor's groundtruth test, which feeds the same
stems through state-actor's streaming builder and geth's UpdateStem +
Commit and diffs the resulting on-disk node sets, fails at n=4 with a
mismatched slot hash in the root group blob.
At a group boundary, recompute the hash via serializeSubtree +
groupedRecursiveHash so that the parent stores the same value the
reader will compute when it deserializes the child blob.
The fix is gated on groupDepth > 0, so nodeStore tests that construct
the store directly without going through NewBinaryTrie retain the
existing pure-SHA256 recursion semantics.
Verification:
- All existing trie/bintrie tests pass unchanged.
- state-actor/generator's TestStreamingMatchesGethCommit (which compares
state-actor's streaming builder output to geth's Commit output
byte-for-byte at n=2,4,8,32,128) now passes.
This PR addresses one of the biggest performance issue with binary
tries: storing each internal node individually bloats the index, the
disk, and triggers a lot of write amplifications. To fix this issue,
this PR serializes groups of nodes together.
Because we are still looking for the ideal group size, the "depth" of
the group tree is made a parameter, but that will be removed in the
future, once the perfect size is known.
This is a rebase of #33658
---------
Co-authored-by: Copilot <copilot@github.com>
The nodes were named using the byte representation of the path, instead
of the binary representation. This was confusing to other client devs
trying to achieve interop.
## Summary
Replace the `BinaryNode` interface with `NodeRef uint32` indices into
typed arena pools, eliminating GC-scanned pointers from binary trie
nodes.
Inspired by [fjl's
observation](https://github.com/ethereum/go-ethereum/pull/34034#issuecomment-4075176446):
> *"if the binary trie produces such a large graph, it should probably
be changed so that the trie node type does not contain pointers. The
runtime does not scan objects that do not contain pointers, so it can
really help with the performance to build it this way."*
### The problem
CPU profiling of the binary trie (EIP-7864) showed **44% of CPU time in
garbage collection**. Each `InternalNode` held two `BinaryNode`
interface values (2 pointer-words each), and the GC scanned every one.
With ~25K `InternalNode`s in memory during block processing, this
created enormous GC pressure.
### The solution
`NodeRef` is a compact `uint32` (2-bit kind tag + 30-bit pool index).
`NodeStore` manages chunked typed pools per node kind:
- **InternalNode pool**: ZERO Go pointers (children are `NodeRef`, hash
is `[32]byte`) → noscan spans
- **HashedNode pool**: ZERO Go pointers → noscan spans
- **StemNode pool**: retains `Values [][]byte` (matching existing
format)
The serialization format is unchanged — flat InternalNode
`[type][leftHash][rightHash]` = 65 bytes.
## Benchmark: Apple M4 Pro (`--benchtime=10s --count=3`, on top of
#34021)
| Metric | Baseline | Arena | Delta |
|--------|----------|-------|-------|
| Approve (Mgas/s) | 374 | 382 | **+2.1%** |
| BalanceOf (Mgas/s) | 885 | 901 | **+1.8%** |
| Approve allocs/op | 775K | **607K** | **-21.7%** |
| BalanceOf allocs/op | 265K | **228K** | **-14.0%** |
## Benchmark: AMD EPYC 48-core (50GB state, execution-specs ERC-20, on
top of #34021 + #34032)
| Benchmark | Baseline | Arena | Delta |
|-----------|----------|-------|-------|
| erc20_approve (write) | 22.4 Mgas/s | **27.0 Mgas/s** | **+20.5%** |
| mixed_sload_sstore | 62.9 Mgas/s | **97.3 Mgas/s** | **+54.7%** |
| erc20_balanceof (read) | 180.8 Mgas/s | 167.6 Mgas/s | -7.3% (cold
cache variance) |
The arena benefit scales with heap size — the EPYC (larger heap, more GC
pressure) shows much larger gains than the M4 Pro (efficient unified
memory). The mixed workload baseline was unstable (62.9 vs 16.3 Mgas/s
between runs due to GC-induced throughput collapse); the arena
eliminates this entirely (95-97 Mgas/s, stable).
## Dependencies
Benchmarked with #34021 (H01 N+1 fix) + #34032 (R14 parallel hashing).
No code dependency — applies independently to master.
All test suites pass (`trie/bintrie` with `-race`, `core/state`,
`triedb/pathdb`, `cmd/geth`).
---------
Co-authored-by: Guillaume Ballet <3272758+gballet@users.noreply.github.com>
