ucx lanes select - sys device distance

[lijh5]

//MD whose distance is queried when evaluating transport selection score
UCX_SELECT_DISTANCE_MD=cuda_cpy

In this function:
static void ucp_worker_iface_set_sys_device_distance(ucp_worker_iface_t *wiface)
{
......
for (i = 0; i < context->num_tls; i++) {
md_index = context->tl_rscs[i].md_index;
if (strcmp(context->tl_mds[md_index].rsc.md_name,
context->config.ext.select_distance_md)) { // Compare with cuda_cpy
continue;
}
device = ucp_worker_iface_get_sys_device(wiface);
cmp_device = context->tl_rscs[i].tl_rsc.sys_device;
ucs_topo_get_distance(device, cmp_device, distance);
}
}

Question:

1. What is the function of UCX_SELECT-DISTANCE-MD and why is it fixed as cuda_cpy?
2. Why is it judged to be equal to cuda_cpy in this function, only then will get distance be executed?
3. Do we not need to get distance for other mds, such as mlx5:0?

Thanks very much!

[yosefe]

UCX_SELECT_DISTANCE_MD=cuda_cpy means that network devices closest to the currently active cuda device have higher score during transport selection. This configuration specifies memory allocation devices and not network devices, so mlx5_0 is not needed here