Drop-In: RocksDB API¶

Same API, same workload, different engine. This benchmark uses identical rocksdb::DB calls -- only the storage backend changes.

What's Being Compared¶

PsiTriRocks: PsiTri's RocksDB-compatible wrapper (libraries/psitrirocks/)
RocksDB: Native RocksDB (linked against system or vendored build)

Both use the same rocks_bench.cpp source compiled against different libraries. The benchmark binary is psitrirocks-bench (PsiTri) and rocksdb-bench (native).

Workloads¶

The standard db_bench workload set:

Workload	Description
`fillrandom`	Insert N random keys
`overwrite`	Overwrite N existing keys (random order)
`readrandom`	Point lookups of N random keys
`readseq`	Sequential scan of all keys
`seekrandom`	Random seeks + short forward scans
`deleteseq`	Sequential deletion of all keys
`deleterandom`	Random deletion of N keys

Parameters: 1M keys, batch size 1, 256-byte values.

Results¶

Benchmark data needed

Run bench/run_all.sh rocksdb_api to collect data on your machine. Requires BUILD_ROCKSDB_BENCH=ON in your CMake configuration.

Results will appear in docs/data/rocksdb_api/<date>/<machine>/.

Bank Benchmark (Transactional Workload)¶

The bank transaction benchmark provides a higher-level comparison using the RocksDB API for realistic transactional operations. PsiTriRocks achieves 95% of native PsiTri performance and 2.9x faster than native RocksDB on TPC-B style workloads:

Engine	Transfers/sec	Relative
PsiTri (native)	376,691	1.00x
PsiTriRocks	356,703	0.95x
RocksDB	126,341	0.34x

See the RocksDB migration guide for API compatibility details and how to switch.

Reproducing¶

# Build with RocksDB comparison enabled
cmake -G Ninja -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_C_COMPILER=clang-20 -DCMAKE_CXX_COMPILER=clang++-20 \
    -DBUILD_ROCKSDB_BENCH=ON -B build/release
cmake --build build/release -j16

# Run the benchmark suite
bench/run_all.sh rocksdb_api

# Or run the CMake comparison target directly
cmake --build build/release --target bench-compare