Drop-In: SQLite API

Same SQL, same schema, different storage engine. This benchmark runs identical sqlite3_* API calls -- only the storage layer changes.

What's Being Compared

• PsiTri-SQLite: SQLite with btree.c replaced by PsiTri's DWAL layer (libraries/psitri-sqlite/)
• System SQLite: Stock SQLite 3.x linked from the system

Both use the same benchmark binary (bank_bench.cpp), the same SQL schema, and the same workload. The only difference is which library is linked.

Architecture

PsiTri-SQLite replaces SQLite's B-tree page storage with PsiTri's DWAL (Dynamic Write-Ahead Log) backed by an ART (adaptive radix trie) buffer. All SQLite tables and indexes share a single DWAL root with 4-byte table ID key prefixes, giving atomic cross-table snapshot consistency.

Application
  |
sqlite3_* API (unchanged)
  |
VDBE / Code Generator / Parser (unchanged)
  |
sqlite3Btree* interface
  |
btree_psitri.cpp  <-- replaces btree.c
  |
PsiTri DWAL root 0 (all tables share one key space)
  [table_id: 4 bytes] [key bytes] → value

Bank Transaction Benchmark

Each transfer performs 5 key-value operations (2 reads + 2 updates + 1 insert) in a single atomic transaction. 10,000 accounts, 100,000 transfers, per-transfer commit (batch=1).

Write Throughput

%%{init: {'theme': 'base', 'themeVariables': {'xyChart': {'plotColorPalette': '#7b1fa2,#1565c0'}}}}%%
xychart-beta
    title "Bank Benchmark — Write-Only (KV ops/sec)"
    x-axis ["PsiTri-SQLite", "System SQLite"]
    y-axis "KV operations per second" 0 --> 500000
    bar [449320, 420765]

Engine	Transfers/sec	KV Ops/sec
PsiTri-SQLite	89,864	449,320
System SQLite	84,153	420,765

Concurrent Read + Write

%%{init: {'theme': 'base', 'themeVariables': {'xyChart': {'plotColorPalette': '#7b1fa2,#1565c0'}}}}%%
xychart-beta
    title "Bank Benchmark — Write + Concurrent Reader (KV ops/sec)"
    x-axis ["PsiTri-SQLite", "System SQLite"]
    y-axis "KV operations per second" 0 --> 400000
    bar [342445, 232330]

Engine	Transfers/sec	KV Ops/sec	Write Impact	Reader reads/sec
PsiTri-SQLite	68,489	342,445	-24%	402K
System SQLite	46,466	232,330	-45%	165K

PsiTri-SQLite delivers 1.47x more write throughput under concurrent reads and 2.4x faster reader throughput. The DWAL's buffered read mode gives readers a consistent snapshot without blocking the writer -- readers never contend with writes.

INSERT Micro-Benchmark

Pure INSERT throughput (100K inserts, no reads, PRAGMA synchronous=OFF):

Mode	PsiTri-SQLite	System SQLite	Ratio
INTKEY autocommit	709K/sec	86K/sec	8.2x
INTKEY batched	1,056K/sec	4,778K/sec	0.22x
TEXT PK autocommit	574K/sec	80K/sec	7.2x
TEXT PK batched	558K/sec	2,124K/sec	0.26x

PsiTri-SQLite dominates autocommit (the realistic per-transaction pattern) because the DWAL's ART buffer makes each commit ~100ns vs SQLite's WAL write overhead. System SQLite wins on large batches because its B-tree insert has no per-op undo log or WAL append.

Tradeoffs

	PsiTri-SQLite	System SQLite
Per-commit cost	~100ns (ART buffer)	~12us (WAL write)
Concurrent readers	Zero writer impact	Blocks during checkpoint
Database format	Directory	Single file
Test suite	83% pass rate	100%
Reader freshness	Bounded by flush interval (1s default)	Immediate

Reproducing

# Build
cmake -G Ninja -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_C_COMPILER=clang-20 -DCMAKE_CXX_COMPILER=clang++-20 \
    -DBUILD_SQLITE=ON -B build/release
cmake --build build/release -j16

# PsiTri-SQLite
./build/release/bin/bank-bench-psitri-sqlite \
    --num-accounts=10000 --num-transactions=100000 --batch-size=1

# System SQLite (for comparison)
./build/release/bin/bank-bench-system-sqlite \
    --num-accounts=10000 --num-transactions=100000 --batch-size=1

# INSERT micro-benchmark
./build/release/libraries/psitri-sqlite/psitri-sqlite-smoke