4 ms·
Not really. It's even 2x slower than SHA256-NI, the builtin. b3sum is much faster than sha256sum, and blake3 is about 2x faster than blake2. http://rurban.git
by rurban 7y ago
Not really. It's even 2x slower than SHA256-NI, the builtin.
b3sum is much faster than sha256sum, and blake3 is about 2x faster than blake2.
http://rurban.github.io/smhasher/doc/table.html http://rurban.github.io/smhasher/doc/table.html
- brandmeyer 7y agoJust exactly what does the "insecure" flag mean in your tables? I'm curious about the kind of definition you are using for security whereby sha2 fails but farmhash passes.
- rurban 7y ago"insecure" means that it fails any of the security tests. SHA2NI and many other hashes fail most of those unless they are hardened, e.g. by support for adding random seeds, being insensitive to added \0 and such. No crypto hash supports random seeds out of the box btw.
- antpls 7y agoNot sure why you are downvoted. Your table looks more factual than the blake3's github.
- espadrine 7y agoRE: the downvotes; while I didn't partake, I know there is history between the BLAKE/BLAKE2/BLAKE3 author, JP Aumasson, and Rurban, in particular related to differing opinions over the publication of SipHash. Differing opinions here means that most of the cryptographic community agree that SipHash clears its claims as a PRF, while Rurban indicated he has found undisclosed vulnerabilities that he wishes not to publish. RE: the table; a point to note is that BLAKE3's main implementation, in Rust, benefits a lot from multithreading on large files, which SHA2 cannot. The C version, which Rurban copied, does not leverage multithreading: https://github.com/BLAKE3-team/BLAKE3/blob/master/c/README.md https://github.com/BLAKE3-team/BLAKE3/blob/master/c/README.m...
- temac 7y agoComparing single threaded speeds is also very valid for tons of contexts. Of course if one algo can't be multithreaded, maybe it will be better to use one that can instead, if that's interesting for a particular application, but it won't be necessarily interesting for all applications. So there is no single ordering about which is faster, in this case. If you sometimes needs to compute hashes on a computer that would have multiple unused cores at that moment, the multithreaded hash will have an advantage. If you are always loaded, you better consider the total amount of work; and maybe likewise if you want to prioritize energy efficiency (although it will not be completely linear)
- zokier 7y agoThe tone of the comment was unnecessarily dismissive, and it is kinda expected that hw implementation beats sw implementations. If anything, it's pretty impressive how close blake3 gets without dedicated instructions.
- kzrdude 7y ago> I.e. the usage of SipHash for their hash table in Python 3.4, ruby, rust, systemd, OpenDNS, Haskell and OpenBSD is pure security theatre. This seems like a noteworthy claim, and the hflounge link is unfortunately broken! Is it possible to you to provide more information about this, that Siphash would be essentially broken?
- rurban 7y agoI never said that Siphash is broken, read what I said. It's not broken, it's just theatre,and way too slow to be useful for anything. And it doesn't protect against hash table attacks. Proper protections need to be done at the collision handling step, not in the hash function. http://perl11.org/blog/seed.html http://perl11.org/blog/seed.html https://news.ycombinator.com/item?id=13361860 https://news.ycombinator.com/item?id=13361860
- zokier 7y agoWhy doesn't sha256sum use sha-ni? On the other hand, seems like sha-ni became widely available only with cannon lake and zen, while blake3 should be fast on any hw even if it benefits from avx512