Scaling advice: watch the connection path in TPS tests
Defaults were used in the benchmark discussed, with one exception: --protocol=prepared. That flag may explain discrepancies if you are seeing much lower throughput numbers.
Another factor is the connection path. On the server, the lighter Unix sockets are used by default for connecting — something impossible for real-life remote clients. That could be a further reason for the gap, and the Unix socket probably should not be used for such TPS tests in the future, since it gives an artificial boost.
