Skip to main content

Benchmarks

DataFusion Nexus benchmark evidence is reported from release builds, with CPU and GPU rows compared from the same harness. The fixture data is generated through fixture/fixture.sh.

Graph Algorithms​

Offical cuGraph benchmarks

TPC-H SF100​

  • Environment: AWS EC2 g7e.4xlarge, 16 vCPUs, 128 GiB system memory, local NVMe fixture storage, RTX PRO 6000.
  • Run: SF100,1 warmup / 1 measured iteration, source cache disabled, LocalKvikio parquet reads.
  • Result: CPU 77.4s, GPU 26.2s, CPU / GPU 3.0x, peak reservation 54.6 GiB.
  • KvikIO: local cuDF/libcudf file reads with cache disabled; this run forced KvikIO compatibility I/O with KVIKIO_COMPAT_MODE=ON. On this g7e.4xlarge host, OFF enters the cuFile path, but cuFile compatibility mode can still fall back; it does not prove native GDS.

Wall time, measured iteration mean.

TPC-DS SF100​

  • Environment: AWS EC2 g7e.4xlarge, 16 vCPUs, 128 GiB system memory, local NVMe fixture storage, RTX PRO 6000.
  • Run: SF100 full suite (99 queries), 1 warmup / 1 measured iteration, profile sf100_gpu_throughput, LocalKvikio parquet reads.
  • Result: CPU 379.4s, GPU 101.7s across 99/99 successful native queries, CPU / GPU 3.7x, peak reservation 43.0 GiB. Row counts matched on 99/99 queries.
  • KvikIO: local cuDF/libcudf file reads with KVIKIO_COMPAT_MODE=ON. On this g7e.4xlarge host, OFF still may enter cuFile compatibility mode without native GDS.

Wall time for the measured iteration. GPU bars use successful native time; failed attempts would still show wall time in gray. Y-axis skips 20–120s so short queries keep readable CPU vs GPU heights; Q72 and other long outliers still render above the break.

Scroll horizontally to inspect all 99 queries.

Full interactive Plotly report: tpcds_sf100_cpu_gpu.html.