Parallel and distributed execution
Execution policy is optional. A local study with no resource constraints does not need an
execution declaration. Adding or changing placement does not change the physical meaning or
study-definition identity.
Bind an execution policy
execution qualification_cluster {
placement = profile.engineering
backend = auto
resources {
workers = 2..64
memory_per_worker <= 32 GiB
accelerators = preferred
}
}
study bracket_response {
geometry = geometry.bracket
mesh = mesh.bracket
model = model.bracket
execution = execution.qualification_cluster
analysis static { structural static {} }
}profile.engineering resolves through project configuration. Cluster addresses, credentials,
trust material, and queue implementation details do not belong in a study file.
backend = auto permits an admitted backend chosen from available capabilities. When a study
requests the accelerator and the accelerator path cannot run it (its solver has no device path,
no accelerator can be allocated, or the device declines the system at solve time), every physics
family runs on the CPU reference path: the result reports cpu_reference as its execution
backend and carries exactly one FEA_ACCELERATOR_ADMISSION_FALLBACK warning, in the published
study as well as on the command line. accelerators = required is the strict option: the study
fails closed with RM.FEA.ACCELERATOR_REQUIRED instead, before any assembly or solve whenever the
fallback is known in advance. The result records the actual backend and device evidence rather
than the preference alone.
Three distinct kinds of scale
RunMat keeps these execution shapes separate:
- Local parallelism runs independent tasks concurrently on one host.
- Cluster fan-out schedules independent sweep cases, frequencies, modes, samples, or recovery partitions on separate workers.
- Cooperative distribution partitions one numerical operator across a worker group with an explicit partition, halo, collective, failure, and checkpoint contract.
The first two do not constitute a distributed solve. Evidence and user-visible status use these terms precisely.
Memory per worker
memory_per_worker <= 16 GiB bounds the resident memory of each worker that runs the study.
Planning works within the effective ceiling: the smallest of this bound, the step's own resident
ceiling, and the memory the host admits. A linear static solid solve selects its path within that
ceiling. When the sparse Cholesky factor fits, the factor proves coercivity and solves. When it
doesn't fit, element rigidity proves the operator positive definite and AMG-preconditioned
conjugate gradients solve it, accepted by the true residual. Each published run records the
effective ceiling (effective_memory_ceiling_bytes) and the selected path (static_solve_path)
in its provenance, and runmat run --json prints both. Neither value enters the
study-definition identity.
Fan-out policy
fanout {
sweep_cases = 16
frequency_samples = auto
}The current typed runner consumes sweep-case fan-out and worker bounds. Frequency fan-out is part of the checked policy surface, but not every procedure currently consumes the authored setting. When a setting is not connected to an executable procedure, planning must not claim that the run used it.
Cooperative policy
cooperative {
operator_partitions = 8
partitioner = graph
balance = nonzeros
}This expresses permission and constraints for procedures with a qualified cooperative execution path. It is not a request to split an arbitrary sparse matrix and hope the solver tolerates it. Admission checks the procedure, formulation, backend, partitioner, communication topology, resource envelope, and worker cohort before dispatch.
Resilience policy
resilience {
retries = 2
checkpoint = required
}Retries resume only from canonical safe boundaries. A retry never changes geometry, material, procedure, convergence, or accuracy intent. Authored resilience fields are checked today, but only procedures with a durable checkpoint implementation may claim resumed numerical work. Otherwise a retry restarts from the latest valid generic boundary or fails admission.
CLI placement
An explicit cluster request uses the same source study:
runmat run bracket.fea \
--study bracket_response \
--cluster engineering \
--project 00000000-0000-0000-0000-000000000000 \
--trust-identity sha256:EXPECTED_IDENTITYEvery explicit --cluster request requires --trust-identity; --project is a UUID. Remote-only
flags are rejected without --cluster. A configured placement profile may supply pinned trust and
routing without these CLI flags. An explicit CLI cluster overrides authored placement but cannot
weaken physical or accuracy requirements.
Accuracy source authority
Portable accuracy tasks for linear structural, modal response-spectrum and random-response results carry the producing workload contract and its canonical numerical workload and mesh inputs. Preparation and workers bind those inputs to the numerical receipt before checking the authored accuracy plan, formulation identity and observation coverage. A result catalog cannot define a weaker acceptance plan or silently omit authored observations. Thermal tasks bind their original thermal workload and mesh directly to the thermal receipt.
The binding is to canonical scientific content, not one transport chunk layout. Rechunking the same payload does not change its producing workload identity. This source validation does not establish convergence, validate a formulation, or prove measured memory bounds. In particular, constraint compilation and observation reprojection still need independently bounded workspace admission; the current workspace multiplier is not a proof for multiplicative expansion.
Constraint-rank classification now borrows the canonical equations and label strings. Its workspace profile bounds coefficient fill and provenance by connected DOF components, with checked slot arithmetic and fallible exact reservations. A count-only graph bound supports admitting connectivity derivation before the tighter numerical workspace profile is available. The same classifier serves compilation and classification validation; validation does not clone the complete rows merely to recompute rank. The profile covers requested Rust heap slots, not allocator overhead, process RSS, diagnostic strings, authored constraint expansion or observation resolution. Whole-task admission must compose those remaining costs before claiming closure.