hwloc
hwloc copied to clipboard
NVML C2C queries for NVIDIA GPU-CPU links
nvmlDeviceGetFieldValues() can retrieve CPU-GPU link info.
#define NVML_FI_DEV_C2C_LINK_COUNT 170 //!< Number of C2C Links present on the device
#define NVML_FI_DEV_C2C_LINK_GET_STATUS 171 //!< C2C Link Status 0=INACTIVE 1=ACTIVE
#define NVML_FI_DEV_C2C_LINK_GET_MAX_BW 172 //!< C2C Link Speed in MBps for active links
(from https://github.com/NVIDIA/nvidia-settings/blob/main/src/nvml.h#L1793)
NCCL already uses those at https://github.com/NVIDIA/nccl/blob/master/src/graph/xml.cc#L702
Those links go to the local CPU (PCI local).