Add numa binding to init from env helper (pytorch#396)

Summary: Pull Request resolved: pytorch#396

Differential Revision: D46036186

fbshipit-source-id: 574307ee09807c9b422239afea643637ce24cb0a

requirements.txt

@@ -2,6 +2,7 @@ torch
 numpy
 fsspec
 tensorboard
+numa
 packaging
 psutil
 pyre_extensions

torchtnt/utils/env.py

@@ -38,6 +38,7 @@ def init_from_env(
     pg_backend: T.Optional[str] = None,
     pg_timeout: timedelta = default_pg_timeout,
     float32_matmul_precision: str = "high",
+    bind_numa: bool = True,
 ) -> torch.device:
     """Utility function that initializes the device and process group, if applicable.
 
@@ -57,6 +58,7 @@ def init_from_env(
         pg_timeout (timedelta, optional): Timeout for operations executed against the process
                                           group. Default value equals 30 minutes
         float32_matmul_precision (str, optional): The setting for torch's precision of matrix multiplications.
+        bind_numa (bool, optional): Whether to bind CPU sockets to GPUs.
 
     Returns:
         The current device.
@@ -86,4 +88,17 @@ def init_from_env(
         )
         torch.distributed.init_process_group(backend=pg_backend, timeout=pg_timeout)
     maybe_enable_tf32(float32_matmul_precision)
+    if bind_numa and device.type == "cuda":
+        init_numa()
+
     return device
+
+
+def init_numa() -> None:
+    import numa
+
+    local_world_size = int(os.environ.get("LOCAL_WORLD_SIZE", 1))
+    num_sockets = numa.get_max_node() + 1
+    socket_id = torch.cuda.current_device() // (max(local_world_size // num_sockets, 1))
+    node_mask = {socket_id}
+    numa.bind(node_mask)