Full duplex communication on a TCP socket

So your send calls are not back to back. I understand in your while loop you actually wait until data is ready from your peripheral and this is correct.

What could be wrong is if you are calling send() every 7us for a single 32bit data.

Nios is a rather slow processor and can't handle such a fast rate. Consider that if you use uC/OS the task switching itself needs a few us. So, you'd better pack more data in a single send call.

With 32bit/7us you need a data throughput about 570kbyte/s: from my experience you can achieve it with Nios running at 50MHz, but you need to optimize the way you handle data. IMHO the SSS sample is good for starting but it has some limitations and must be adapted to the actual application.