This benchmark measures the affect of the number of queue pairs on sender throughput. Each server thread connects to many clients.
This benchmark is for a single NIC. If dual_port == 1,
each thread (client or server) uses port id % 2. In this case, server threads
connect to clients whose ID is the same as the server's modulo 2.
For READs, we limit the number of outstanding
operations per-thread to WINDOW_SIZE. This is because too many outstanding
requests can cause WQE cache misses, and this benchmark focuses on QP cache
misses.
A similar restriction is harder for WRITEs because we cannot detect completion
by polling on memory. For WRITEs, we have a reasonable limit on the number of
outstanding operations per-QP (i.e., around 2 * UNSIG_BATCH), not per-thread.
The per-thread limit is around 2 * UNSIG_BATCH * NUM_CLIENTS.
The rw-tput-sender benchmark should be used to measure outbound performance
with low QP fanout.