Struct RpcAgent 

pub struct RpcAgent { /* private fields */ }

RPC agent that manages function registration and remote invocation.

Each rank creates an RpcAgent wrapping a Backend. Functions are registered with [register] and invoked remotely with [rpc_sync].

Response correlation

Concurrent rpc_sync calls are correlated by request_id. If a received response has a different request_id than expected, it is buffered and the agent retries the recv. Buffered responses are checked before issuing new recv calls.

Implementations

impl RpcAgent

pub fn new(backend: Arc<dyn Backend>) -> Self

Create a new RPC agent.

pub fn register<F>(&self, name: &str, handler: F)

where F: Fn(&[u8]) -> Result<Vec<u8>, String> + Send + Sync + 'static,

Register a callable function.

The handler receives serialized arguments and must return serialized results. If the registry lock is poisoned, this recovers the inner data and continues.

pub fn rpc_sync( &self, dst_rank: usize, function_name: &str, args: &[u8], ) -> FerrotorchResult<Vec<u8>>

Invoke a function on a remote rank synchronously.

Sends the request, then waits for a response with the matching request_id. If a response for a different request is received, it is buffered for later retrieval.

Errors

Returns an error if the remote function is not found, if the remote handler returns an error, or if communication fails.

pub fn rpc_async( self: &Arc<Self>, dst_rank: usize, function_name: &str, args: &[u8], ) -> JoinHandle<FerrotorchResult<Vec<u8>>>

Invoke a function on a remote rank asynchronously.

Spawns a thread to perform the RPC call. Returns a join handle that can be used to retrieve the result.

Limitations

Spawns an unbounded number of OS threads. Each call to rpc_async creates a new thread. This is acceptable for infrequent coordination RPCs but is not suitable for high-frequency patterns. A thread pool or async runtime would be needed for that use case.

pub fn handle_request( &self, src_rank: usize, request_data: &[u8], ) -> FerrotorchResult<()>

Handle an incoming RPC request: look up the function, call it, and send the response back.

pub fn rank(&self) -> usize

The rank of this agent.

pub fn world_size(&self) -> usize

The world size.