I'm a big believer in benchmarking before saying things, but I think we could skip the benchmarks when asking the question "does 1 4096-byte read or 4096 1-byte reads complete faster".
I would agree ... except that I've seen cases in which it didn't make a difference.
e.g. if the File implementation had an internal buffer (C stdio's "FILE " does), and the read from that* buffer was inlined (from my past experience up to date as of early 2011, HotSpot doesn't, but LuaJIT does), it might not make any difference.
Seriously, LuaJIT does things I've never thought I'd see a compiler (JIT or AOT) for any language (dynamic or statically typed) do. I used to reply to "sufficiently smart compiler" with "one hasn't appeared yet, despite at least 3 decades of waiting". But LuaJIT has appeared.