gigl.distributed.graph_store.compute#

Attributes#

R

logger

Functions#

async_request_server(server_rank, func, *args, **kwargs)

Perform an asynchronous request on a remote server, calling on the client side.

init_compute_process(local_rank, cluster_info[, ...])

Initializes distributed setup for a compute node in a Graph Store cluster.

request_server(server_rank, func, *args, **kwargs)

Perform a synchronous request on a remote server, calling on the client side.

shutdown_compute_proccess()

Shuts down the distributed setup for a compute node in a Graph Store cluster.

Module Contents#

gigl.distributed.graph_store.compute.async_request_server(server_rank, func, *args, **kwargs)[source]#

Perform an asynchronous request on a remote server, calling on the client side.

Parameters:

  • server_rank (int)

  • func (Callable[Ellipsis, R])

  • args (Any)

  • kwargs (Any)

Return type:

torch.futures.Future[R]

gigl.distributed.graph_store.compute.init_compute_process(local_rank, cluster_info, compute_world_backend=None)[source]#

Initializes distributed setup for a compute node in a Graph Store cluster.

Should be called once per compute process (e.g. one per process per compute node, once per cluster_info.compute_cluster_world_size)

Parameters:

  • local_rank (int) – The local (process) rank on the compute node.

  • cluster_info (GraphStoreInfo) – The cluster information.

  • compute_world_backend (Optional[str]) – The backend for the compute Torch Distributed process group.

Raises:

ValueError – If the process group is already initialized.

Return type:

None

gigl.distributed.graph_store.compute.request_server(server_rank, func, *args, **kwargs)[source]#

Perform a synchronous request on a remote server, calling on the client side.

Parameters:

  • server_rank (int)

  • func (Callable[Ellipsis, R])

  • args (Any)

  • kwargs (Any)

Return type:

R

gigl.distributed.graph_store.compute.shutdown_compute_proccess()[source]#

Shuts down the distributed setup for a compute node in a Graph Store cluster.

Should be called once per compute process (e.g. one per process per compute node, once per cluster_info.compute_cluster_world_size)

Parameters:

None

Return type:

None

gigl.distributed.graph_store.compute.R[source]#
gigl.distributed.graph_store.compute.logger[source]#