Repository navigation
Demonstrator
Oguzhan KAYNAK edited this page Nov 14, 2026
·
1 revision
One program that uses both planes. Gray-Scott reaction-diffusion is a common test application of in situ analysis research; here it is a vehicle for the model, not a goal.
The full walkthrough — what is computed, how the work is split, each phase with diagrams, how ranks synchronise — is in apps/gray_scott/README.md. The numerical specification is docs/gray-scott.md.
- Coordinator process: queries versions, starts the job, sends heartbeats, follows progress, collects results, shuts down. Control plane only.
- Worker processes: a main thread on the control plane, worker threads on the data plane.
- Workers form a simulation group and an analysis group.
| Phase | Work | Pattern |
|---|---|---|
| Startup | version query, job | control plane |
| 1 Distribute | initial field to the simulation group | scatter |
| 2 Simulate | steps; edge rows to neighbours every step | neighbour exchange |
| 3 Measure | every K steps: ΣU, ΣV | allreduce |
| 4 In transit | every K steps: V rows from simulation to analysis | M × N redistribution with acknowledgements |
| 5 Analyse | histogram of V, combined in pairs, then pairs of pairs | reduction tree |
| 6 Collect | final fields to rank 0, results written | gather, barrier |
| Throughout | heartbeats, progress | control plane |
| Check | Expectation |
|---|---|
| Distributed run vs single-threaded run | bit-identical, for every split |
| C++ vs Python (numpy) reference | within a tolerance (relative 1e-9) |
| Deployment over TCP vs over shared memory | byte-identical fields |
| ASan, UBSan, TSan | clean |
Dynamic load balancing, fault tolerance, groups defined by configuration, security. Groups are fixed in code and phases are separated by barriers.
gridlane
Repositories