I'm not sure the saving is much bigger. I would expect that in most cases if the two files differ most of the blocks will be different (e.g. add a byte to a fil

I'm not sure the saving is much bigger. I would expect that in most cases if the two files differ most of the blocks will be different (e.g. add a byte to a file at the beginning, now no block is the same). I think that block-level deduplication has more predictable performance though: for example if you write some blocks to a deduplicated file, with block-level deduplication you just write the new blocks, with file-level deduplication you have to rewrite the whole file.