21G on tc egress is slightly surprising to me. I'd like to see the program used for the benchmark. Was GSO accounted for? If you pop/pull headers by hand, you'll often kill GSO which will result in a massive loss in throughput like this.