Tony Allen

Are you running it from the Sia-GPU-Miner directory?

I'm all for introducing a more sophisticated load balancing technique, but we should be really understand _why_ what we have today does not address your specific use-case. Before getting into...

> The reason is that when the source pod chooses a destination pod, the active_request queues are mostly empty or at a size of 1/2, so the algorithm works as...

This is great. Thanks for the extra information. So, the `cx_active` stats are showing us the number of active **_connections_**. Envoy's LEAST_REQUEST load balancer normalizes endpoint weights by the number...

Thanks for the extra graphs, that helps in seeing the differences between backends. > We think that by applying different weights based on historical response time, the spread of the...

> I think that success latency + error rate can provide a good measurement for weight. I'm not opposed to the idea, we just need to be more specific on...

@jizhuozhi can you just briefly describe how all of this works here? I don't know how mosn works and would appreciate it if you could clearly describe how this new...

> First, the LEAST_REQUEST load balancer (LRLB) calculates the expected completion time of all tasks based on the number of active requests. It assumes that all servers have the same...

> What I need to clarify is that I understand how Least Request is implemented in envoy, but what I want to express is that usually Least Request is abstracted...

That's some convincing data! I'll need to do a more in-depth read, but from my quick pass the data in section 5.2 makes a good case for implementing the **Prequal**...