Helper function to parse size strings like "1K", "2M", "3G"

Note: post() must return NIXL_IN_PROG in case of success

Post the slot's request and record post_duration. Marks the slot in-flight; the caller drives completion via getXferStatus.

Register memory (if --reregister_mem) and create the XferReq for a slot that doesn't already have one. Records the wall-clock time as prepare_duration

Helper to prepare transfer descriptors based on backend type

* @brief Print usage information for the program * @param program_name The name of the program executable */

Run the nixlbench command with appropriate parameters. This method builds a command for the nixlbench tool, including only n

Per-rail completion processing - handles one rail's CQ with configurable blocking behavior

Query information about memory/storage # Arguments `descs` - Registration descriptor list to query `opt_args` - Optional arguments specifying backend

Return the ranks that receive messages

Register a memory buffer with the agent

Initiate remote memory transfer.

Poll /metrics until it contains `needle`, or timeout.

Send a notification to a remote agent # Arguments `remote_agent` - Name of the remote agent to send notification to `message` - The notification mess

* @brief Set metadata (only valid for fragment 0) * @param total_payload_length Total length of combined payload across all fragments * @par

* @brief Set payload chunk for this fragment using move semantics * @param payload Chunk of (agent_name + message) combined payload (passed by va

Set the batch size for benchmarking.

Set whether notification is enabled

Set the notification message

Records `key`'s current value once, so Drop can restore it even when the key was not passed to `new`.

Return the sum of the sizes received by <rank>

Wait for all asynchronous memory queries to complete with a per-descriptor wait timeout.

Generates a random number in [0,255] (byte range). */

Configure GDS and GDS_MT plugins (same logic for both)

Configure OBJ plugin for object storage operations.

Configure POSIX and HF3FS plugins (same logic for both)

Configure UCX, GPUNETIO, and Mooncake plugins (same logic for all)

Find the first phase where this rank appears after the start_search_from_phase

Returns lists of buffers where bufs[i] is the send/recv buffer for rank i

Get the id for a group of ranks

Get the next available ID and increment the counter

Return the CUDA major version that torch was built for, or None.

Return the CUDA major version that torch was built for, or None.

Map only the header to read buffer size

Map the file into memory

Open existing file

Set default values for parameters that were not explicitly provided. This method is called during initialization to ensure all requi

Share receive buffer descriptors between all ranks, in alltoall style. Args: my_recv_bufs: List of receive buffers for current ran