Seems to me like the 10 GbE performance on the LX2160A RDB is pretty poor. A simple client/server test moving large blocks of data between the LX2160A RDB and another computer using a TCP socket and send/recv produces max speeds of about 4.8 Gb/sec for sends and about 6 Gb/sec for recv. The exact same code, with the exact same network connections produces about 7.8 Gb/sec between two Centos 7.5 machines (one of which is used as the other end point for the LX2160A RDB test) and 8.7 Gb/sec between two machines running Solaris 11.
One adjustment we've found that really helped performance is interrupt coalescing. There doesn't seem to be a way to control that in the dpaa2 driver. Does anyone know if that can be done?
Also, is there some other driver we should be using for the LX2160A 10-GbE?
Are there other adjustments that have a big effect on throughput - we're just using the defaults for everything else (except we are using jumbo frames of 9000 bytes). We just haven't gone any further because no matter what those interrupts will hammer the CPU.
