From my testing, the auto weighing has memory, going from idle to saturation has no ramp up delay. Unlike MLVPN, reordering buffer is working correctly with up to 200ms difference between links. Its missing link quality monitoring, and/or FEC. It uses hardcoded values as of now like loss_tolerance, the value seems to be very tolerant unfortunately as it's linked with other logic. The lack of link monitoring causes issues like when one link latency is very high, it will slow down the entire tunnel instead of DC. There is no bufferbloat when using Ubond, even with just one link that exhibits bufferbloat directly. High bandwidth asymmetry between links doesn't weigh the slowest connection to 0 it will keep using it. I followed MLVPN documentation.