Add the ability to specify frame alignment

[asayers]

This PR is a proof-of-concept which adds a --align=<n> flag which forces all frames to be aligned to an n-byte boundary in the resulting file. It does this by inserting skippable frames as padding.

Demo

Frame Index     Compressed    Uncompressed    Compressed Offset    Uncompressed Offset 
          0    1008.69 KiB        5.75 MiB                  0 B                    0 B                 
          1      15.31 KiB             0 B          1008.69 KiB               5.75 MiB            
          2    1015.99 KiB        3.75 MiB             1.00 MiB               5.75 MiB            
          3       8.01 KiB             0 B             1.99 MiB               9.50 MiB            
          4     998.34 KiB        3.62 MiB             2.00 MiB               9.50 MiB            

This is the result of running with --align=1M. The odd-numbered frames are the regular ones (the ones you'd want to read from). Notice how they all start on 1 MiB boundaries. The even-numbered frames are skippable "padding" frames.

Motivation

In theory, setting --align=4K should reduce read amplification and pagecache usage when seeking. The effect would be more noticeable with smaller frames. Eg. if you wanted a perfectly seekable file, you'd want 4 KiB frames. In that case, misaligned frames would be amplified by 100%. I haven't measured anything though, so I don't really know how big the effect is in practice.

[asayers]

I wouldn't be surprised if there are bugs in this implementation. This PR is meant to be more of a conversation-starter than something that can be merged as-is.

[rorosen]

Very interesting, thanks.

Do you have a use case for aligned frames? I like the idea but suspect that it will be rarely useful in practice

[asayers]

The idea is based on the fact that files are not really random-access: you can only really seek to a multiple of 4k. So if the frame you want isn't 4k-aligned then the kernel will end up reading some useless data from the end of the previous frame. This doesn't matter much if you seek rarely, so this --align is meant to be a performance optimisation for the case where you do a lot of small random reads. That's the theory, but I still haven't benchmarked anything, so I still have no idea if it's effective!