Sorting String is much faster than sorting Vec<u8>

[vigna]

I noticed that sorting String values is much faster than sorting the same values represented a Vec<u8>. I expected the second case to be faster. Is there an easy explanation?

[dapper91]

@vigna Hi. Could you provide your test? According to my measurements u8 typed items are sorted faster than String.

[vigna]

First, thank you for the prompt answer! You must be a little patient with me because I have a lot of experience in other languages but this is, well, my first Rust program.

It takes a directory with a number of large TSVs (the motivation was BlockChair files with cryptocurrency transitions), cuts some fields and sorts the result. There is an u8-vector (src/bin/u8.rs) version and a string (src/bin/string.rs) version. The u8 version uses ByteLines to read directly vectors of bytes. https://github.com/vigna/blockchair-sort/

You run it by passing the directory where the files are present and the list of fields. For example,

./target/release/u8 testoutputs 2 5

where testoutputs contains TSVs.

The u8 program is eight times slower than the string program on the same data and I'm really puzzled.

I have flamegraphs https://vigna.di.unimi.it/string.svg and https://vigna.di.unimi.it/u8.svg showing what happens on my hardware. I'm not that familiar with the inner workings of Rust but it looks like in the u8 case there is a lot of serialization/deserialization.