Those numbers seem awfully low; even going via the relatively slow kernel/kernel stack, you can comfortably achieve around 2Mpps (64B) tx, per core (single flow

Those numbers seem awfully low; even going via the relatively slow kernel/kernel stack, you can comfortably achieve around 2Mpps (64B) tx, per core (single flow), on ~2010 commodity hardware via standard Linux userspace APIs. For reference, bypassing the kernel you can saturate a 10G link for ~14Mpps on a single downclocked 500MHz core with same class of hardware. Stuck on phone at the moment; will check out the code later.