Monte Carlo is often embarrassingly parallel, while optimisation may mix independent candidates with coupled linear algebra. This note selects the simplest efficient execution layer and keeps stochastic evidence reproducible.
Correctness baseline
Implement a small deterministic or fixed-seed reference and verify statistical outputs, confidence intervals and constraints. Assign non-overlapping seed streams by scenario/worker and record the mapping. Reproducibility means equivalent statistics and traceable random inputs, not always bitwise identity across libraries.
Single-node scale
AxiKeystone provides 128 cores and 1.15TB RAM for process pools and in-memory datasets. Measure one worker, then increase concurrency while tracking memory bandwidth, serialisation and NUMA locality. Batch short tasks so scheduling overhead does not dominate.
GPU path
AxiFoundry provides a professional 24GB GPU for CUDA-enabled Python/Julia/MATLAB experiments. RAPIDS accelerates supported tabular pipelines. Keep arrays on device across steps and include transfer time in comparisons. Irregular branch-heavy simulations may remain better on CPU.
Two-node path
Use AxiVector when the dataset exceeds one node, the solver exposes MPI or independent work can be distributed with low coordination. Run fixed-problem and growing-problem tests over 100GbE RDMA. Do not accept more workers if efficiency collapses or licence cost outweighs saved time.
Campaign operation
Checkpoint optimiser state and completed scenario IDs. Use a scheduler for fair allocation and failure retry. Report wall time, scenarios/second, parallel efficiency, peak RAM/VRAM, convergence and uncertainty. Archive environment, code, inputs, seeds and selected outputs.
Primary technical references
- RAPIDS API documentation
- Dask-CUDA documentation
- Slurm MPI Users Guide
- NVIDIA CUDA C++ Best Practices Guide
References are provided for software architecture and implementation planning. Validate the versions, licences, support matrix and regulated-use requirements applicable to the final deployment.