[ofa-general] Re: [PATCH] mlx4: increase max outstanding rdma reads per qp

[Roland Dreier]

 > Change max outstanding rdma reads per QP from 4 to 16.
 > This enables an improvement in latency for rdma-read applications.

This only affects performance if an app queues more than 4 RDMA READ
requests, right?  (Because the 5th request doesn't have to wait for
the 1st request to complete before it's sent)

Do we want to increase this for mthca too, or is mlx4 so much faster
than mthca that we need more requests in flight to keep the pipeline
full?

 - R.