Show HN: Go version of HighwayHash with optimized assembly implementations

Show HN: Go version of HighwayHash with optimized assembly implementations(github.com)

48 points by y4m4b4 8 years ago | 13 comments

loeg 8 years ago |

How does performance compare with XXhash, possibly the current leader in the non-cryptographic hash category?

rurban 8 years ago | |

XXhash is only fast with digest workloads, not hash tables. There the hash function needs to be fast and small, XXhash is too big and trashes the icache.

The current leaders are on my smhasher site, https://github.com/rurban/smhasher/

loeg 8 years ago | | |

Thank you!

zuzun 8 years ago | |

The benchmarks on the xxHash Github page were done on a 32-bit system using an old Visual C++ compiler. Of course you can beat hash functions that were explicitly designed for modern 64-bit architectures. The Core 2 lacks the CRC instruction introduced by the SSE 4.2 extension, which is the main reason Google wrote their hashes.

loeg 8 years ago | | |

Look at XXH64.

bla2 8 years ago | | |

xxHash has a 64-bit version.

danesparza 8 years ago |

This seems incredibly fast. Taking the benchmarks on the sites at face value, this seems like it's roughly twice as fast as xxHash.

I also appreciate this line in the README: "HighwayHash is not a general purpose cryptographic hash function (such as Blake2b, SHA-3 or SHA-2) and should not be used if strong collision resistance is required"

I am curious to know what the threshold for "strong collision resistance" is.

rurban 8 years ago | |

It's used for 32bit hashes in hash tables, so brute forcing this to create thousands of collisions on a desktop PC takes 2-4 min. Same for SipHash btw.

For hash tables you should not rely on the false security claims of these slow hashes anyway, rather implement proper collisions strategies to get rid of O(n) attacks.

fwessels 8 years ago | |

Check out the graph at https://github.com/fwessels/HashCompare that compares different hashing techniques.

And also we've done some predictions for collisions, see: https://github.com/fwessels/HashCompare/issues/1#issuecommen...

gpderetta 8 years ago | |

Not a cryptographer, but basically it must not be computability feasible (ie not just hard) to find a collision.

tromp 8 years ago |

Note that performance is strongly dependent on message size. My Cuckoo Cycle proof of work hashes roughly a billion 8-byte messages with the 'cryptographically strong' SipHash in 4 seconds, for a single-core speed of 2000MB/s, while HighwayHash on longer 16-byte messages slows down to only 200MB/s (no figure provided for 8-byte messages).