> I suggest we do this in mthca too. Have you tested this to know whether it matters? Increasing the limit uses more memory per QP... Does the rdma read latency test in OFED queue up enough work requests to measure this? - R.