Struct MpiExecutionGroup 

pub struct MpiExecutionGroup { /* private fields */ }

One application-owned MPI communicator duplicated for Eqiora admissions.

Implementations

§

impl MpiExecutionGroup

Signature links:MpiExecutionGroup

pub fn duplicate<C>( communicator: &C, provided: Threading, required: MpiThreadSupport, ) -> Result<MpiExecutionGroup, Diagnostic>

where C: Communicator,

Signature links:duplicate&CMpiThreadSupportResultMpiExecutionGroupDiagnostic

Duplicate an initialized communicator and validate thread support.

MPI initialization/finalization remains entirely with the application. provided must be the level returned by initialize_with_threading.

Errors

Returns EQ0807 if rank/size do not fit the Eqiora partition contract or the provided threading level is below required.

pub const fn partitions(&self) -> NonZero<usize>

Signature links:partitionsNonZerousize

Number of ranks/partitions in this execution group.

pub const fn partition(&self) -> PartitionId

Signature links:partitionPartitionId

Local rank represented as an Eqiora partition identity.

pub const fn thread_support(&self) -> MpiThreadSupport

Signature links:thread_supportMpiThreadSupport

Thread-support level actually provided at MPI initialization.

pub fn solver_capabilities(&self) -> SolverCapabilities

Signature links:solver_capabilitiesSolverCapabilities

Exact numerical policies admitted by the current distributed evidence slice.

pub fn agree_rank_device_topology( &mut self, local: RankLocalDeviceV1, ) -> Result<MpiRankDeviceTopologyV1, Diagnostic>

Signature links:agree_rank_device_topologyRankLocalDeviceV1ResultMpiRankDeviceTopologyV1Diagnostic

Collectively bind one distinct physical device to every MPI rank.

V1 requires the launcher to expose exactly one device to each process, making the local ordinal zero on every rank. The fixed-width physical identity prevents two ranks on the same host from silently selecting the same device. The returned rank-ordered evidence and fingerprint are identical on every rank.

Errors

Returns a common EQ0807 diagnostic when an ordinal is not zero, a physical identity is zero or duplicated, allocation fails, or the gathered record count contradicts the execution group.

pub fn agree_composed_local_readiness( &mut self, local: &Result<(), Diagnostic>, ) -> Result<(), Diagnostic>

Signature links:agree_composed_local_readinessResult()DiagnosticResult()Diagnostic

Agree readiness of fallible rank-local work in a composed adapter.

A composed adapter calls this before entering its next MPI collective: after local capture/allocation/session preparation and again after any fallible post-execution evidence validation or summary construction. This prevents a ready rank from entering a collective while another rank returns from local work. No payload or backend-specific identity crosses this seam.

Errors

Returns a common admission-phase diagnostic when any rank reports local readiness failure.

pub fn agree_composed_execution_summary( &mut self, local_summary: [u8; 32], ) -> Result<(), Diagnostic>

Signature links:agree_composed_execution_summaryu832Result()Diagnostic

Require one fixed-width composed-execution summary on every rank.

A composition adapter calls this after the ordinary MPI execution receipt has been independently agreed. The adapter’s summary can bind topology, transport, local-action counts, and that receipt without exposing a general byte registry or MPI communicator. This seam applies its own domain separation before comparison.

Errors

Returns a common EQ0802 diagnostic when rank summaries differ.

pub fn admit<'group, 'model>( &'group mut self, system: &'model DistributedLinearSystem, complete: &'model CanonicalCsrSystemView, plan: SolverPlan, ) -> Result<AdmittedDistributedRun<'group, 'model>, Diagnostic>

Signature links:admitDistributedLinearSystemCanonicalCsrSystemViewSolverPlanResultAdmittedDistributedRunDiagnostic

Collectively seal one system, complete verifier, and sole solver plan.

All dynamically sized communication and verification storage is fallibly reserved before the first admission record is exchanged. A scalar readiness reduction makes allocation/validation failure common before any rank can enter the record gather.

Errors

Returns the same stable diagnostic on every participating rank for a system, layout, plan, verifier, count, or workspace contradiction.

Trait Implementations

§

impl Debug for MpiExecutionGroup

Signature links:DebugMpiExecutionGroup

§

fn fmt(&self, formatter: &mut Formatter<'_>) -> Result<(), Error>

Signature links:fmtFormatterResult()Error

Formats the value using the given formatter.

§

impl MpiAdmittedExecutionAdapter for MpiExecutionGroup

Signature links:MpiAdmittedExecutionAdapterMpiExecutionGroup

§

fn execute_admitted( &mut self, admitted: AdmittedExecution<'_>, ) -> Result<AcceptedLinearExecution, Diagnostic>

Signature links:execute_admittedResultDiagnostic

Execute one graph-bound distributed solve and return its common L2 receipt with a transport-normalized actual collective trace.

§

fn execute_admitted_with_local_action( &mut self, admitted: AdmittedExecution<'_>, action: &mut dyn MpiRankLocalCsrAction, ) -> Result<AcceptedLinearExecution, Diagnostic>

Signature links:execute_admitted_with_local_actionMpiRankLocalCsrActionResultDiagnostic

Execute one graph-bound distributed solve while delegating only the rank-local owned-row CSR action.

§

impl MpiCudaAdmittedExecutionAdapter for MpiExecutionGroup

Signature links:MpiCudaAdmittedExecutionAdapterMpiExecutionGroup

§

fn execute_admitted_mpi_cuda( &mut self, admitted: AdmittedExecution<'_>, ) -> Result<AcceptedMpiCudaLinearExecution, Diagnostic>

Signature links:execute_admitted_mpi_cudaResultAcceptedMpiCudaLinearExecutionDiagnostic

Execute host-staged MPI Krylov with one resident CUDA CSR action on every rank.