I applied the attached patches on top of your patch, so it would be nice to see if you can give it a try in your test hardware to see whether it helps.
I can confirm that I no longer see any regression at 8 or even 16 clients. In fact, the patched version consistent do better than master even at higher number of clients.
The only thing I am a bit puzzled is that I am no longer able to reproduce the 40-50% gains that I'd earlier observed with a single client. I now get 20-25% with smaller number of client and 10-15% with larger number of clients. I haven't been able to establish why it's happening, but since it's a different AWS instance (though of the same type), I am inclined to blame it on that for now.
