When NICs are overwhelmed with egress traffic (and return ENOBUFS when attempting to enqueue traffic), lacp metadata may be dropped or severely delayed, causing lagg link flapping.
Mellanox NICs allocate per-send-tag transmit queues, so if we allocate a rate-limited send tag for lacp, lacp gets to avoid the congested TX queues and send its metadata on an uncongested queue. This allows lacp metadata to be delivered to the link partner, avoiding link flaps when under very high load.
This approach was suggested by Mellanox. I'm not at all sure it is a good idea. My concern is that for nics like Chelsio, which support rate limit, but do not support allocating a separate queue for each rate limited send tag, this may actually delay lacp metadata even more, if it is prioritized behind non-rate limited traffic.