Correct me if I'm wrong, but this code would block if (for example) one single request in a batch of 10 takes significantly longer to complete. The 9 other threads are then doing nothing while they wait.

Of course, you could implement a more robust algorithm for scheduling threads, but it sounds like an implementation of FiberScheduler would make the best use of a single thread's resources (and you could scale that to multiple threads as needed using something like a shared queue). Since fibers are lighter-weight, it's presumably more efficient to run parallel IO operations in a single thread, and only use more threads if your work requires it.