mars

HPC Parallel Execution Guide

This guide converts the HPC contracts into a parallel implementation plan for subagents. It is operational guidance; docs/hpc_contracts.md remains the contract source of truth.

Recommended Start Order

1. hpc_contract_governance_20260511
2. H0 packaging cleanup and validation tracks:
   • spack_upstream_submission_20260511
   • easybuild_upstream_submission_20260511
   • conda_forge_feedstock_submission_20260511
3. hpc_cpu_parallel_runtime_20260511
4. hpc_abi_arrow_runtime_boundary_20260511
5. hpc_accelerator_portability_20260511
6. hpc_distributed_execution_20260511 (CPU cluster parallelism and partitioned replay)
7. hpsf_e4s_readiness_submission_20260511
8. julia_general_registration_submission_20260511

H0 packaging tracks can run in parallel with H1 benchmark-baseline work. H2, H3, and H4 should not start implementation until H1 has produced stable semantics and benchmark evidence.

Parallel Subagent Slices

Agent	Track	Write scope	Blocking dependencies
A	Governance + claims	docs/hpc_contracts.md, claim-check script/docs, release/community references	None
B	H0 packaging lane	packaging/spack/**, packaging/easybuild/**, packaging/conda-forge/**, submission notes	Contract governance baseline
C	H1 runtime lane	Rust runtime kernels, Rust/Python benchmark and parity tests	Contract governance baseline
D	H2 ABI lane	ABI/boundary files and tests, interoperability notes	H1 semantics and benchmark thresholds
E	H3 accelerator lane	Optional array-module replay backend and docs/tests	H1 benchmark data
F	H4 distributed lane	CPU cluster parallelism, command-backed worker dispatch, and distributed adapter files and tests	H1 partitioning semantics; H3 only if accelerator-distributed

Parallel Team Assignment (6 Subagents)

The request for six parallel workers maps cleanly to one-lane ownership:

• Agent A: Governance lane (hpc_contract_governance_20260511) and claim-check enforcement.
• Agent B: H0 packaging lanes (spack, easybuild, conda-forge) and any blockers from external tooling.
• Agent C: H1 runtime optimization, benchmark evidence, and serial/parity tests.
• Agent D: H2 boundary hardening, versioning contract, and boundary conformance.
• Agent E: H3 optional replay, optional dependency/feature-gate design, and safety policy.
• Agent F: H4 CPU cluster parallelism, command-backed worker dispatch, distributed partitioning, and deterministic aggregation work.

Notes:

• HPSF/E4S and Julia registration tracks are still handled as governance-facing follow-ups after the primary lanes above have clear checkpoint evidence.
• The H4 lane includes the replay-only local preview adapter and an explicit command-backed worker adapter. The command adapter is available only when a worker command is configured and does not imply hidden scheduler provisioning.

Coordination Rules

• Each agent must state its write scope before editing.
• Agents must not revert or rewrite files owned by another lane.
• Cross-lane API changes must be proposed in the owning track before edits.
• H0 agents must remove non-final URLs, versions, checksums, and stale package identities before upstream-bound files are considered complete.
• H1 agents must define benchmark thresholds before changing kernel behavior.
• H3 agents must keep vendor accelerator claims separate from the current array-module validation backend unless vendor parity and benchmark evidence has been added.
• HPSF/E4S agents must not submit full packets unless H0 is complete and H1/H2 evidence is credible; otherwise they should draft pre-submission inquiries or deferral notes.
• Julia registry agents should keep their artifacts review-only and avoid touching runtime/kernel files while preparing Registrator materials.

Validation Commands

Baseline repository checks:

jq . docs/release_metadata.json
rg -n "HPC-ready|HPSF|E4S" docs conductor packaging
./scripts/check_hpc_claims.sh
cat docs/hpc_claim_review_checklist.md

Track artifact checks:

for d in conductor/tracks/hpc_*_20260511 conductor/tracks/*submission_20260511; do
  test -f "$d/spec.md"
  test -f "$d/plan.md"
  test -f "$d/metadata.json"
  test -f "$d/index.md"
  jq . "$d/metadata.json" >/dev/null
done

Benchmark evidence checks:

cd rust-runtime
cargo bench --bench runtime_bench -- --noplot
python3 scripts/benchmark_runtime_threads.py --help

This site is open source. Improve this page.