Bench: native Rust vs GraphAr C++ vs graphar-rs
Three harnesses take the same commands and print the same JSON lines:
| dir | implementation | what it covers |
rust/ |
this crate (graphar-bench) |
prepare inputs, info files, write (table and row-builder API), read a property, read CSR, full scan |
cpp/ |
GraphAr C++ library (gar_bench_cpp) |
write (VertexPropertyWriter/EdgeChunkWriter and builder::*), the same reads with the chunk readers |
graphar-rs/ |
upstream graphar-rs 0.1.0 (graphar-rs-bench) |
vertex row builder only (it has no edge builder and no readers yet) |
Every reader runs on the graphs written by every table writer, and on the upstream testing/ldbc_sample graph (written by the Spark library), so the run doubles as an interoperability check: the check object of a read (row counts, CSR sums, column sums) must be the same for every reader and writer.
Inputs
An input dir holds vertices.parquet, edges.parquet (with _graphArSrcIndex, _graphArDstIndex and the edge properties) and schema.json.
graphar-bench prep-gar data/ldbc-input testing/ldbc_sample/parquet/ldbc_sample.graph.yml person person knows person
graphar-bench prep-mimir data/mimir-input <dump-dir>... # rustc graph dumps (Arrow IPC)
for f in parquet csv orc arrow; do graphar-bench info data/ldbc-input $f data/ldbc-info-$f; doneBuild
All builds run in the container of Containerfile (Ubuntu 24.04, Arrow C++ 24.0.0 from the Apache Arrow apt repository, Rust 1.91.1).
podman build -t graphar-bench -f Containerfile .
# GraphAr C++ (apache/incubator-graphar checkout at $UP)
cmake -S $UP/cpp -B build/cpp -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$PWD/install
cmake --build build/cpp && cmake --install build/cpp
cmake -S cpp -B build/bench-cpp -G Ninja -DCMAKE_PREFIX_PATH=$PWD/install && cmake --build build/bench-cpp
# this crate
(cd rust && cargo build --release --features orc)
# graphar-rs: upstream is a symlink to the checkout
ln -s $UP graphar-rs/upstream && (cd graphar-rs && GRAPHAR_BUILD_TYPE=Release cargo build --release)Run
mkdir -p work/bin && cp <the three binaries> work/bin/ && cp -r data work/
LD_LIBRARY_PATH=$PWD/install/lib UPSTREAM_TESTING=$UP/testing RUNS=5 ./run.sh $PWD/work <label>
python3 summarize.py work/raw/<label>.jsonlrun.sh writes one JSON line per run with the peak RSS of the process (/usr/bin/time -f %M); secs is the time of the operation itself, measured inside the process (input loading for the writers is not counted).