@Crypt-iQ you’re completely right, thanks for highlighting this: I did mean TCP packet and assumed the best-case where it isn’t fragmented by intermediate routers. But it was part of my open-ended question about possible performance degradation: I have no idea whether in practice 65kB TCP packets often get fragmented or not, how much overhead it adds, and couldn’t find public data about this…if a lot of fragmentation happens, then padding could indeed lead to degraded performance (on top of the additional bandwidth usage increase).
Is this a default OS behavior? Or is it just a recommendation? I’ll read the links you provided when I have some time, but would love a TL;DR for now 
I have no idea where to find actual data of what happens on the internet nowadays. I was thinking that we could do A/B testing on mainnet nodes: pad all packets to 65kB for a few days, then remove padding for a few days, and repeat, while measuring latency. This could give some indication of the overhead, even though it wouldn’t let us accurately predict higher percentiles since there probably isn’t enough payment volume today to provide meaningful statistics, but it’s a start.