Using the newer sys_epoll system call with Knot avoids this
  problem and achieves excellent scalability. However, we 
  have used the poll() result for comparison, since sys_epoll
  is incompatible with Haboob's socket library.

WTH? I admit I didn't read the entire thing (just skimmed), but are they actually comparing threads to poll() and declaring victory instead of epoll()? Intentionally? epoll() is radically different from poll(). Nobody uses poll() for heavy-duty event-based servers. They might as well not have said anything at all.