I've modified the source to generate timesamples (as a seperate stream) from the USRP2 source blocks and noticed an issue - if the PC is struggling some frames have a timesample dated before the timesample of a frame processed before it.
It's possible I've done something wrong modifying the code but I think this is a host issue as the problem is reduced (i.e it only happens at lower decimation rates) on faster machines.
Has this been seen before? Are there any kernel options etc that might improve this? I'd assumed the ethernet bit was a FIFO buffer of some sort but guess this isnt working quite right.