Performance and memory

RustSTA separates timing propagation from mapped result queries. Measure the workflow you actually need: input import, graph lowering, analysis, persistence, and any host/device synchronization all contribute to elapsed time.

This documentation release publishes no numerical RustSTA-vs-OpenSTA speedup. The source contains benchmark harnesses; a harness is not itself a measured result. Avoid treating a GPU propagation microbenchmark as end-to-end timing performance.

Measure an analysis

free -h
(
  ulimit -v 16777216
  RUSTSTA_PHASES=1 /usr/bin/time -v -o timing.time timeout 180s \
    ruststa run --workspace ./ws --base-rev rev1 \
      --threads 2 --max-paths 10 --max-state-gib 4 --format json \
      > timing.json 2> timing.stderr
)

Use a fresh, equivalently prepared workspace for each mutating repetition. Record cold/warm cache policy, host CPU, worker allocation, input hashes, backend, precision, corner, wire model and unconstrained endpoint count. Report WNS/TNS, path/check agreement and peak RSS beside wall time. If import is outside the timer, label the number analysis-only.

The RUSTSTA_PHASES=1 diagnostics help locate import, graph and propagation costs. /usr/bin/time -v supplies whole-process maximum resident set size. Neither replaces a complete command boundary measurement.

Bound the workload

Estimate state as objects × bytes per entry × corners × workers, accounting for per-node/arc vectors, timing tags and worker scratch. Ranked path counts do not bound every other state table. Keep --max-paths and --max-state-gib explicit when running unfamiliar inputs.

Use topological propagation and bounded extraction rather than enumerating all paths through reconvergent logic. Restrict report output with --limit. Run large verification jobs with two to four test workers and serialize jobs estimated above 4 GiB using flock .work/scale.lock.

For CUDA, prefer a cgroup RSS/swap guard so virtual address reservations work:

systemd-run --user --scope -p MemoryMax=16G -p MemorySwapMax=4G \
  timeout 180s ruststa run --workspace ./ws --base-rev rev1 \
    --backend cuda --precision f64 --max-paths 10 --format json

If the user service manager is unavailable, arrange an equivalent OS/container memory limit before a large GPU run. Do not apply a CPU address-space cap to a CUDA process.

Existing harnesses

The CLI package includes run_e2e and backends benchmark targets, and the PDK suite includes large store/query cases. Run them only after an input-specific memory estimate and host preflight. Compare CUDA results to CPU reference results and external timing to pinned OpenSTA outputs.