I've updated the benchmarks to include those flags. I've also specified the versions of all software involved.
It seems par2 is significantly faster with those options set than without, as in by an order of magnitude, it seems par2 struggles greatly with the large number of blocks that it sets by default. Thank you for telling me.
haha! Everything in the source code is prefixed with bef_ to avoid any possible symbol collisions when linking, so it'd be a pain to unlearn prefixing it.
>and even 10000 and it still worked!!!! pretty awesome
Mind you, bursts greater than a certain size, around 8000 bytes for the defaults, are a game of probability in what damage they will do. In most cases it'll work like you did for you, but in some unlucky offsets 10000 byte corruptions with default settings could do something like this.
1 byte corrupts fragment x of block n, 4096 bytes decimate fragment x of block n+1, 4096 bytes decimate fragment x of block n+2, 1807 bytes corrupts fragment x+1 of block n.
The defaults only have 1 parity fragment, so the fact two fragments were corrupted for block n means it can't be faithfully reconstructed. In the average case you should be fine, but it is something I should probably communicate more effectively in the README (I only hint at it with the 8K per 192K comment).
Got it, I'll add par2cmdline-turbo (I didn't know it existed) to the list along with those key details. Likewise, I'll also go and describe the benefits and downsides of each tool in more detail. I'll get around to it when I release the next version soon that fixes some of the problems described in this thread.
>Is that because, as you increase the interleave, it increases the amount of the stream you need to read in order to repair any corruption?
Yes, as you need to read the entire set of interleaved blocks to get each respective block, so to keep memory consumption low, we don't want to interleave too many blocks. I could increase the default to something higher though.
Regarding the tape problem, I was being a bit daft in my response, as if you lose a chunk of tape, you also lose some indeterminate number of data with it, which my format currently isn't capable of recovering from. I'll try to fix that to some degree in the next version, at the cost of not guaranteeing backwards compatibility as I feel it falls under that major problem comment I made in the v0.1 release.
> i bef'ed the Makefile, removed a character, tried to 'deconstruct' it, the output is zero bytes
I can't reproduce this. These are the commands I used, with it freshly compiled on a ubuntu docker container, both the v0.1 release and the master tree.
edit: oh, I see, you 'removed a character'. Depending on what character you removed or corrupted from the output, you could've either hit the issue described above with inserting noise, but this time removing information, or you caused corruption in the header by either corrupting the magic number, hash type, or hash itself. The command line utility automatically truncates the output before calling the deconstruct function. The header is sadly the biggest single point of failure in the tool/format, which is why I introduced the --raw flag for those who don't want it.
the parity fragments are by default just stored at the end, interleaved like every other fragment. It doesn't actually matter whether a parity or a data fragment get corrupted, as long as there are at least some determinate number of them not corrupted.
In terms of controlling how many blocks and their respective fragments are interleaved, that can be controlled by the -l argument in the command. I'll paste the info I have in the manpage here.
>The number of blocks(specifically their fragments) to interleave. The default number is 3, as it provides protection for a bad burst to corrupt both the block in front of and behind it.
In general, you can approximate the burst size ratio needed to destroy a set of interleaved blocks beyond repair via this equation (this may not be fully accurate as its napkin math)
let n = number of blocks to interleave, B = number of bytes per fragment, m = number of parity fragments, k = number of required fragments to rebuild a block.
((n - 1) * m * B + n - 1) / (n * (k + m) * B)
this is because a given corruption could traverse a whole fragment and then leech into another, taking some other fragments with it with the most unlucky of bursts. When you remove the B and take the limit as n goes to infinity, it approaches the ratio of m/(k+m), or in other words as you interleave more and more blocks, you get a maximal burst size closer and closer to the size of your total number of parity fragments.
Also, the header specifies exactly the size of each fragment, we don't depend on the fragments themselves to tell us their size. In fact, the only metadata a fragment has is its hash and the number of padded bytes for the whole interleaved set of fragments, the format is (almost, ignoring padded bytes) completely described from the header as a design principle. That's why a whole fragment can be turned to garbage and we don't care, just skipping ahead to the next fragment.
edit: thinking on it, I could add a mode of behavior such that it recovers as much as possible if a specific block is corrupted, rather than the situation right now where it quits streaming after it finds a corrupted block.
Sadly my tool currently doesn't account for that type of corruption, as it doesn't know what is good data or bad data when reading. So if bad data is inserted between symbols/fragments, rather than corrupting the symbols/fragments themselves, the tool would read them naively and exit with an error when the hashes don't add up. I'm sure there's a clever way of defending against that, but as of this moment I'm not entirely certain how best to do so.
Currently there is just one set of defaults, but I could very well add multiple different defaults dedicated to certain use cases such as the ones you have. I imagine it'd be something like this
I was not, as I am a english speaker with no knowledge of Dutch. I find it a funny coincidence, but if I change the name bef now it'll mess with my muscle memory.
I'm not familiar with zbackup, but from a google search appears to be a tool to deduplicate and encrypt data. The process envisioned while making this was to use a series of pipes unix style to make the backup, e.g.
tar c dir | zstd | gpg -e | bef -c -o backup.tar.zst.gpg.bef
and then to get back that file with the terribly long filename
bef -d -i backup.tar.zst.gpg.bef | unzstd | gpg -d | tar x
Sure, erasure coding is a form of error correcting codes that can be applied to data such that you can lose some n number of codes before you can successfully reconstruct the input data. For example, take k input symbols, and put it into an erasure code algorithm to get k+n symbols, where any n of the output symbols can be lost before you fail to be able to reconstruct the data. Symbols in this case can be some number of bits/bytes.
This is a really important property in situations where there can be big giant bursts of errors, because you can still reconstruct the data regardless. IIRC, CDs/DVDs/BDs all use two concatenated Reed Solomon (a type of erasure coding) coded symbols that are then interleaved with each other, which provides the disk protection against things like accidental scratches.
Hello Hacker News! I'm been developing this piece of software for about a week now, to serve as a fast and easy to use replacement for par2cmdline and zfec. Now that it is in good and presentable state, I'm releasing it to the world to get users, feedback, testing on architectures that aren't x86[-64], etc. If you have any feedback, questions, or find any bugs/problems, do let me know.
It seems par2 is significantly faster with those options set than without, as in by an order of magnitude, it seems par2 struggles greatly with the large number of blocks that it sets by default. Thank you for telling me.