PDES and Distributed
KairoECS has implemented conservative and optimistic parallel discrete-event simulation in Rust, with MPI and typed gRPC transport paths. Their evidence is stronger than a scaffold-only description, but it has explicit boundaries.
Current evidence
Section titled “Current evidence”| Area | What the evidence supports | What remains separate |
|---|---|---|
| Conservative PDES | Single-host event processing, final-state parity and repeated wall-clock measurements for the recorded ring workload. | General speedup, multi-host behaviour and useful scaling on other model families. |
| Optimistic Time Warp | Bounded rollback, antimessage, GVT, durable recovery and entity-transfer qualification. | A proof for all models, all failure schedules or arbitrary deployments. |
| Native MPI | Real two/four-rank smoke and bounded distributed execution cases. | Cross-host MPI, network-fabric behaviour, large-rank and cluster-scheduler acceptance. |
| gRPC | Native Mac/Colab and local transport cases in retained qualification receipts. | General WAN reliability, production security posture and end-to-end throughput. |
PR #366 was merged on 3 November 2026. Its native-runtime qualification and associated receipts describe exact execution scope. Draft/unmerged wording in historical captures refers to their capture time.
Performance is an open engineering task
Section titled “Performance is an open engineering task”The measured PDES benchmark
retains five samples per configuration on an Apple M1 Max at source
45f01c49be1d15cefb64ad48e59ee0a7e4146b3e. For those small ring workloads,
sequential throughput was about 4.4–5.0 million events/s while PDES achieved
about 6.7–23 thousand events/s. Each round created and joined OS workers, and
the 16/32-LP cases exceeded the host CPU count.
These are historical measurements, not a current-revision benchmark or a comparison with another simulator. Use sequential execution as a measured baseline, profile coordination cost, and evaluate persistent workers and larger useful work per synchronization round before recommending parallelism. See Track 111.
Generated scaling fixtures provide no independent scaling or parity evidence. Preserve exact discrete replay, floating-point tolerance and statistical-equivalence claims separately.