Struct DistributedOptimizationContext 

pub struct DistributedOptimizationContext<M: MPIInterface> { /* private fields */ }

Distributed optimization context

Implementations

impl<M: MPIInterface> DistributedOptimizationContext<M>

pub fn new(mpi: M, config: DistributedConfig) -> Self

Create a new distributed optimization context

pub fn rank(&self) -> i32

Get the MPI rank of this process

pub fn size(&self) -> i32

Get the total number of MPI processes

pub fn is_master(&self) -> bool

Check if this is the master process

pub fn distribute_work(&mut self, total_work: usize) -> WorkAssignment

Distribute work among processes

pub fn synchronize(&self) -> ScirsResult<()>

Synchronize all processes

pub fn broadcast_parameters(&self, params: &mut Array1<f64>) -> ScirsResult<()>

Broadcast parameters from master to all workers

pub fn gather_results( &self, local_result: &Array1<f64>, ) -> ScirsResult<Option<Array2<f64>>>

Gather results from all workers to master

pub fn allreduce_sum( &self, local_data: &Array1<f64>, ) -> ScirsResult<Array1<f64>>

Perform all-reduce operation (sum)

pub fn allreduce_min( &self, local_data: &Array1<f64>, ) -> ScirsResult<Array1<f64>>

Perform all-reduce operation (element-wise minimum)

pub fn broadcast_from( &self, params: &mut Array1<f64>, root: i32, ) -> ScirsResult<()>

Broadcast a vector from an arbitrary root rank to all processes

pub fn select_global_best( &self, local_best: &Array1<f64>, local_value: f64, ) -> ScirsResult<(Array1<f64>, f64)>

Select the globally-best candidate across the whole communicator.

Given each process’s local best point and its objective value, returns the best (point, value) pair over all processes. The global-best value is obtained with a Min all-reduce; the owning rank (lowest rank on ties) is identified by a second Min all-reduce over rank indices, and its point is broadcast so every process agrees on the winning candidate.

pub fn ring_exchange( &self, outgoing: &Array1<f64>, ) -> ScirsResult<Option<Array1<f64>>>

Exchange a vector around a unidirectional process ring.

Sends outgoing to rank (rank + 1) % size and returns the vector received from rank (rank - 1 + size) % size. A deadlock-free ordering is used (rank 0 receives before sending, every other rank sends before receiving), so it is safe with blocking point-to-point primitives for any communicator size. Returns None when there is only a single process.

pub fn stats(&self) -> &DistributedStats

Get performance statistics