We implemented a Gigabit Ethernet capability for the BF561 (running uClinux, kernel 2.6.x @ 600MHz) using an ASIX AX88180 chip. Once everything was in place we learned that we could only obtain a fraction of the gigabit performance, namely just bellow 100Mb/s. After getting 4kB jumbo frames to work the performance improved slightly to ~114Mb/s. I measured that it takes ~102us to push the 4kB packet into the ASIX chip (at kernel network driver level, "xmit" function). Yet from the throughput (tested with iperf) it looks like the system requires a total of ~270us to send a 4kB packet.
At the moment I am wondering if:
- it is expected that the layers above the driver network need ~170us per 4kB frame?
- using kernel 3.x would help?
- there any internal scheduling policies that I might try to optimize?
- if there is someone that obtained gigabit performance/capability using VisualDSP++?
Thank you for your time and help.