Five execution variants over identical generated directed weighted graphs: Icebug (C++/Arrow NetworKit update), Icecat (Rust/Arrow rewrite), Grustcat (Grust/Arrow), Grustcat Cypher (Grust parser and typed Arrow backend), and official Neo4j GDS.
Published benchmark: https://adversari.al/graph/algorithms. The graph index also links the separate graph-query and strain benchmarks.
git clone https://github.com/querygraph/adversarial-graph-algorithms
cd adversarial-graph-algorithms
./docker/run.sh --frozen --full-path --algorithms dijkstra --families path \
--sizes 16384 65536 --warmups 0 --repeats 1 --label reproduced-full-path--frozen selects the published measured snapshot. Without it, run.sh measures
current upstream Grust and Turso with Arrow, DataFusion and mimalloc; see below.
Requires Docker Compose v2 and Python 3.12+ (safe tar extraction). docker/prepare.py defaults to the checksum-verified measured source snapshot in publication/; no sibling repositories are required. Official Neo4j/GDS archives are downloaded and checksum-verified separately. For local development, python3 docker/prepare.py --local stages ../icecat and ../grust; then invoke Compose directly. The snapshot includes the exact staged build inputs, while transitive OS packages/base-image tags are not immutable. Separate dependency licenses are under publication/licenses/; bundled sources retain their own notices.
The 65,536-node chain produces 2,147,516,416 entries in each path array per engine. Neo4j has no transaction deadline in full-path mode. One measured run and no warmup demonstrate completion and correctness; they do not establish stable performance rankings. GDS includes server-side Cypher overhead. Grustcat Cypher includes parsing, semantic analysis, planning, typed Arrow execution and aggregation; it is not the general Grust reference executor.
Protocol and resource limits · raw evidence and checksums · source snapshot receipt.
All five variants also passed 30 small cases (six families × five algorithms); the runtime image passes 45 kernel/Arrow interchange checks. BFS and PageRank have documented cross-engine output/convergence differences. Graph loading and projection are outside kernel/query timers. No published query-suite result is pooled with these measurements.
Generalized graph algorithms in Grust describes the proposed upstream architecture, Arrow and Cypher contracts, phased acceptance gates, and benchmark validation. It includes committed source references and clone commands for an implementation agent on another machine.
./docker/run.sh with no source flags measures current upstream Grust and Turso:
the historical four participants plus upstream direct and Cypher, both Turso
snapshot columns, Arrow and DataFusion, built with the mimalloc global allocator.
Checkouts come from docker/upstream-pins.json; a missing checkout is cloned at
its pinned commit, and a checkout sitting at any other commit is refused rather
than silently measured. Pass --upstream-grust or --turso to measure a
different source on purpose, which is recorded as unpinned in the run receipt.
The current Grust/Turso workflow adds upstream direct/Cypher execution, Arrow/DataFusion preparation, and Turso-backed snapshot participants, with process-wide mimalloc by default and explicit allocator and release-profile controls. Turso main is pinned by commit for each experiment. The historical reproduction command above continues to use the published frozen sources. Group commit is measured in a separate concurrent-write preparation workload, outside algorithm timers.
Related work places these participants against the best-known Rust graph algorithm library and states what has not been measured.
The current-source report carries the measured results, boundaries, regressions and evidence, kept separate from the historical publication. The write-up covers the output contract and what measuring it on current sources found.