6M! that's great. But that's not comparing apples to apples, the best test they did was using a 40 core machine with a 10gbe NIC.
These cards usually have as many receive/send queues as the number of cores the machine has, while the ec2 instances only have 2 queues. (and most performance gains in packet receive usually come from handling interrupts on multiple CPUs...)
Would be great to see results for ulib on c3.8xlarge machines.