WIP perf changes from Garand's branch#5238

Draft

dmkozh wants to merge 75 commits intostellar:masterfrom

dmkozh:garand_perf_test

Commits on Feb 18, 2026

Fix devcontainer to work with noble

graydon
authored and
Garand Tyson
committed
Rewrite copilot-instructions, add some skills.

graydon
authored and
Garand Tyson
committed
Tracy analysis skill

SirTyson
authored and
Garand Tyson
committed
Add subsystem-summary skills

graydon
authored and
Garand Tyson
committed
Enable parallel ledger apply for apply-load max-sac-tps mode
Garand Tyson
committed
Add max-sac-tps optimization skills and enable in-memory BucketList
Garand Tyson
committed
Improve binary search for apply-load.

dmkozh
authored and
Garand Tyson
committed
Speed up benchmark: tighter search range, fewer samples, add ralph loop setup
Garand Tyson
committed

Commits on Feb 19, 2026

apply-load: Speed up noisy binary search for max-sac-tps benchmark

SirTyson
and
claude
committed
Fix CXXSTDLIB default and standardize build config for clang-20/libc++
Garand Tyson
committed
Shard verifySig cache to reduce mutex contention (7680→8896 TPS, +15.8%)
Garand Tyson
committed
Optimize commitChangesToLedgerTxn with WithoutLoading APIs (8896→9408 TPS, +5.8%)
Garand Tyson
committed
docs: document failed experiment 008 - budget charge() fast path optimization
Garand Tyson
committed
Ralph iteration 1: work in progress
Garand Tyson
committed
Add experiment 009 failure doc: cache init entry XDR (+1.4%, marginal)
Garand Tyson
committed
Parallelize addLiveBatch and InMemorySorobanState update in commit path (saves ~56ms/ledger)

Garand Tyson
and
claude
committed
Parallelize InMemoryIndex construction with bucket put loop (saves ~25ms/ledger)

Garand Tyson
and
claude
committed

Commits on Feb 20, 2026

Cache TransactionFrame::getSize() to avoid redundant xdr_size computation (saves ~500ms/trace)

Garand Tyson
and
claude
committed
Switch SHA256 from libsodium (pure C) to OpenSSL (SHA-NI hardware accel)

Garand Tyson
and
claude
committed
perf: overlap per-thread commit with parallel execution (+13.6% TPS)

Garand Tyson
and
claude
committed
perf: disable BUILD_TESTS meta tracking in benchmark (-50ms/ledger, -3.4%)

Garand Tyson
and
claude
committed
perf: eliminate per-tx child LTX in fee processing (+19.2% TPS)

Garand Tyson
and
claude
committed
perf: skip invariant delta when no invariants enabled (+8.0% TPS)
Garand Tyson
committed
perf: indirect sort in convertToBucketEntry (+2.8% TPS)
Garand Tyson
committed
Ralph iteration 1: work in progress
Garand Tyson
committed
Add large binary/profiling artifacts to .gitignore

Garand Tyson
and
claude
committed
Ralph iteration 1: work in progress
Garand Tyson
committed
Ralph iteration 2: work in progress
Garand Tyson
committed
Ralph iteration 3: work in progress
Garand Tyson
committed
Ralph iteration 1: work in progress
Garand Tyson
committed

Commits on Feb 21, 2026

perf: skip encoded_key XDR serialization in Soroban apply path

Garand Tyson
and
claude
committed
perf: disable budget metering for max-sac-tps benchmark

Garand Tyson
and
claude
committed
perf: cache initial entry XDR sizes to skip old-entry re-serialization

Garand Tyson
and
claude
committed
perf: eliminate initial storage map clone in production

Garand Tyson
and
claude
committed
perf: cache Budget via thread-local storage across TXs

Garand Tyson
and
claude
committed
perf: add Tracy zones to invoke_host_function sub-operations

Garand Tyson
and
claude
committed
perf: add Tracy zones to try_finish and event externalization

Garand Tyson
and
claude
committed
perf: add Tracy zones to try_finish and event externalization

Garand Tyson
and
claude
committed
perf: skip redundant OperationFrame::checkValid in preParallelApply

Garand Tyson
and
claude
committed
perf: skip redundant validation in commonValidPreSeqNum during apply

Garand Tyson
and
claude
committed
perf: replace std::set with UnorderedSet for eviction modified keys

Garand Tyson
and
claude
committed
perf: eliminate child LTX in refundSorobanFee

Garand Tyson
and
claude
committed
perf: skip child LTX in removeAccountSigner via peek

Garand Tyson
and
claude
committed

Commits on Feb 22, 2026

perf: skip checkOperationSignatures for Soroban TXs

Garand Tyson
and
claude
committed

Commits on Feb 23, 2026

perf: remove dead loadFromLedger call and add diagnostic Tracy zones

Garand Tyson
and
claude
committed
perf: skip child LTX in preParallelApply when meta is disabled

Garand Tyson
and
claude
committed
perf: remove ZoneScoped from trivial hot functions

Garand Tyson
and
claude
committed
perf: share LedgerTxnHeader in processRefund and move result pair

Garand Tyson
and
claude
committed
perf: track entry existence in ParallelApplyEntry to skip SHA256 lookups

Garand Tyson
and
claude
committed
perf: move semantics in commitChangesToLedgerTxn to avoid XDR copies

Garand Tyson
and
claude
committed
perf: skip child LTX in processFeesSeqNums when meta disabled

Garand Tyson
and
claude
committed
perf: pre-load Soroban RO entries + processFeesSeqNums optimizations

Garand Tyson
and
claude
committed
perf: eliminate UnorderedSet from recordStorageChanges

Garand Tyson
and
claude
committed
fix: revert p21-p24 submodules to upstream and push p25 to SirTyson fork

Garand Tyson
and
claude
committed
fix: revert p26 submodule to upstream commit

Garand Tyson
and
claude
committed
fix: regenerate protocol-25 ledger close meta golden files

SirTyson
and
claude
committed
fix: three correctness bugs causing test failures

SirTyson
and
claude
committed

Commits on Feb 24, 2026

Added overview of experiment
Garand Tyson
committed
perf: cache getTTLKey computation in ThreadParallelApplyLedgerState

Garand Tyson
and
claude
committed
perf: in-place mutation in InMemorySorobanState avoids erase+emplace

Garand Tyson
and
claude
committed
perf: reserve parallel apply container capacity to eliminate rehashing

Garand Tyson
and
claude
committed
perf: cache CxxLedgerInfo cost params serialization per ledger

Garand Tyson
and
claude
committed
perf: cache TTL key lookups in addReads using thread-local cache

Garand Tyson
and
claude
committed
perf: cache TTL key hash in InternalContractDataMapEntry ValueEntry

Garand Tyson
and
claude
committed
perf: move entries instead of copying in getAllEntries

Garand Tyson
and
claude
committed
perf: avoid building 128K-entry modifiedKeys set for eviction scan

Garand Tyson
and
claude
committed
Update readme with last experiment
SirTyson
committed
Add missing docs
Garand Tyson
committed
Update readme again
SirTyson
committed
perf: replace InMemoryBucketEntry virtual set with unordered_map
Garand Tyson
committed
Experiment 068: Fuse auth-check + balance-load in SAC transfer (+1.3%)
Garand Tyson
committed

Commits on Feb 25, 2026

Experiment 075: Remove high-frequency Rust Tracy zones (+1.0%)
Garand Tyson
committed
Exp 078: Use upsertEntryKnownExisting for RW entries known to exist in recordStorageChanges
Garand Tyson
committed
perf: add Tracy zones to unzoned Rust invoke functions for profiling visibility
Garand Tyson
committed

Commits on Apr 20, 2026

apply load changes
dmkozh
committed