Skip to content

Execution-resource benchmarks and security exploits

Summary

Adds cores×threads-per-core competitive sweep CSV, SIMD intrathread labeling, lic resource flags (--cores, --threads-per-core), and execution_exploits CI gates for oversubscribe and memory-bomb caps.

Agent continuation

  1. Read benchmarks/harness/execution_resource_sweep.py, benchmarks/competitive/registry.toml (execution_resource_sweep track).
  2. Run ./scripts/build.sh then python3 benchmarks/harness/execution_resource_sweep.py --smoke and ./li-tests/run_all.sh execution_exploits.
  3. Next: stack PR on #206; full sweep on Linux CI (bench.py --tier 6).
  4. Blocked: none.

Changed

  • compiler/common/resource_options.{hpp,cpp} — --cores, --threads-per-core, cap host cores, effective_jobs memory cap
  • compiler/lic/main.cpp — wire flags on check/build
  • benchmarks/harness/execution_resource_sweep.py, bench.py --tier 6, bench_toolchain.py --resource-sweep
  • benchmarks/runtime_refs/reduce_parallel/ — li_native, cpp_omp, cpp_pthread
  • li-tests/execution_exploits/run.sh, scripts/ci-security.sh, run_all.sh
  • benchmarks/competitive/registry.toml — execution_resource_sweep ecosystem

Not changed

  • Tier-0 energy/invariant rows (bench.py --tier 0 / stability.py thresholds)
  • Li parallel codegen for reduce (li_native uses serial ref kernel)

Breaking

N/A — additive harness and CLI flags.

Security

  • execution_exploits: --cores=999999 capped, negative --threads/--cores rejected, --jobs memory bomb capped via effective_jobs

Performance

  • benchmarks/results/execution_resource_sweep.csv — wall_s, peak_rss_mb, speedup, efficiency = speedup/(cores×tpc)

Downstream

  • li-langverse/benchmarks dashboard may ingest execution_resource_sweep.csv when published