Struct Collectives 

pub struct Collectives<D: CollectiveDriver> { /* private fields */ }

The groups this device has joined.

One per server, unlike the identifiers behind group_id: an identifier is what every rank of a group agrees on, a communicator is one rank’s membership of it.

Implementations

impl<D: CollectiveDriver> Collectives<D>

pub fn new(device: DeviceId) -> Self

A device that has joined nothing yet.

pub fn mark_broken(&mut self, group: &CommunicationId, cause: ServerError)

Records that this device’s part of a collective over group failed with cause. Its peers wait in that collective, and another one over the group would pair with theirs.

pub fn ensure_intact(&self, group: &CommunicationId) -> Result<(), ServerError>

Refuses a collective over group once a part of one here has failed.

Errors

ServerError::Generic naming the failure that broke the group.

pub fn join( &mut self, devices: Vec<DeviceId>, ) -> Result<Option<CommunicationId>, ServerError>

Join the group over devices, and answer with the id it is known by.

A group already joined is joined once: Ok(None) says there was nothing to do, so the caller does not announce a membership twice.

Errors

ServerError::Generic when this device is not among devices — it would have no rank in a group it is not in — and whatever the driver says about agreeing an identifier or joining.

pub fn transfers_with( &mut self, peer: DeviceId, ) -> Result<&D::Communicator, ServerError>

The communicator for transfers between this device and peer, joined on first use.

Errors

Whatever the driver says about agreeing an identifier or joining.

pub fn get( &self, devices: &CommunicationId, ) -> Result<&D::Communicator, ServerError>

This device’s membership of the group over devices.

Errors

ServerError::Generic when this device never joined that group, which is a missing join rather than anything the device did.

pub fn rank( &self, device: DeviceId, devices: &[DeviceId], ) -> Result<usize, ServerError>

The rank of device in the group over devices.

Errors

ServerError::Generic when device is not among devices.

pub fn peer_rank(&self, devices: &[DeviceId]) -> Result<usize, ServerError>

The rank of the one device in devices that is not this one.

For the two-device operations — a send has exactly one peer, and its rank is whichever position this device does not occupy.

Errors

ServerError::Generic when every device in the pair is this one, so there is no peer to name.

Trait Implementations

impl<D: CollectiveDriver> Debug for Collectives<D>

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

Formats the value using the given formatter.