Ettus USRP B200 full duplex at 16 or 32 MSPS?

Question

I am testing the B200 with two different PCs, both running Ubuntu 14.04 LTS.
One is a 4 year old Core i7 PC with an Intel MB and a NEC Corporation uPD720200 USB 3.0 controller.
The other is a 1 year old Asus G750 laptop with an Intel 8 Series/C220 USB 3.0 controller. 
Both machines can handle sustained (a few hours) 350+ MB/second transfers from a Point Grey USB 3.0 camera over a 3 meter cable with zero USB 3.0 errors reported by Point Grey's software. Note this is one direction of data flow over USB 3.0, not full duplex. The PC reached 356 MB/second. The laptop reached 369 MB/second. Both error free over periods of about 1 hour. 
The B200 can reliably TX and RX (full duplex) at a common data rate of 4 MSPS on both machines. 
8 MSPS full duplex is intermittently OK (occasional buffer overruns on both machines). Note: (8 MSPS) * (2 bytes/sample) * 2 (RX + TX) * 2 (I + Q) = 64 MB/second... fairly low for what USB 3.0 is capable of.
16 and 32 MSPS full duplex is never reliable (all runs result in buffer overruns and missing RX or TX data).
Command I'm using (set for 8 MSPS):
nice -20 ./txrx_loopback_to_file --file samples.dat --type float
--nsamps 50000000 --tx-rate 8000000 --rx-rate 8000000 --tx-freq
915000000.000000 --rx-freq 916000000.000000 --tx-gain 0.000000
--rx-gain 0.000000 --wave-type RAMP --wave-freq 10000.000000
Note this is 1 channel TX + 1 channel RX full duplex. Wire format is "sc16" (16 bits per sample over the bus I believe). Data format is "fc32" (32 bit float). I've tried both the stable and unstable UHD drivers. 
I also observed that a moderate bend of the USB 3.0 cable will noticeably reduce reliability. Nothing too severe as to permanently deform or damage the cable. 
I suspect RFI is affecting the USB 3.0 signals. 
*** Questions ***
Any suggestions on how to achieve reliable 16 or 32 MSPS full duplex transfers? Hopefully reliable as in sustained full-duplex TX and RX or an average of 15 minutes between data transfer errors. 
Has anyone placed their USRP B200 or B210 in a metal enclosure? Was reliabliity affected?
Has anyone tested their B200 / B210 with different USB host controllers and found a controller that is ideal for full-duplex TX+RX at the 16 or 32 MSPS rates? 
Do full duplex TX+RX configurations create a significantly greater I/O challenge over USB versus TX-only or RX-only?
Or is the USRP B200 / B210 simply not the best hardware for doing sustained 16 or 32 MSPS full duplex TX and RX? Better to use a PCI Express interface SDR?

Answer 1

John,
Thank you for your time and help.
Yes I have posted to the USRP-users mailing list along with more detailed test results. I've also corresponded with Matt Ettus.
Thank you for the link to benchmarks. I later found that. I have stopped using the B200 with the computer with the NEC controller although it did give similar results versus the laptop with the 8 Series/C220. I am now just testing with the Asus Laptop with the 8 Series/C220 controller.
I do appreciate that the PCIe or 10-gigabit based SDRs are more reliable. Matt told me this too. I would expect this based on the prices. I am more concerned that full duplex with the B200 appears almost useless. Maybe my 8 Series/C220 controller has electrical or RFI problems? Maybe my B200 is defective?
I am concerned with the $5000 cost for a PCIe or 10-gigabit Ethernet solution when a PCIe x4 capable FPGA now only cost $100 or $200 more than the FPGA and Cypress FX3 that the B200 uses.
I'm tempted to instead spend $5K to roll my own PCIe MIMO SDR for $750 to $1000 per board versus spending $5000 for one X300, RF daugtherboard and external PCIe kit. And I can customize it to my needs better. Per Vices did exactly this and offered a $750 (later revised to $2500) PCIe SDR board with 100 MHz simultaneous TX and RX bandwidth.
Jason