Lessons from Hash Table Merging
gist.github.com
Lessons from Hash Table Merging
1–10 of 23 posts
Re: Lessons from Hash Table Merging
#2Kudos, neat digging and writeup that makes us think :)
If you merge linear probed tables by iterating in sorted hash order then you are matching the storage order and can congest particular parts of the table and cause the linear probing worse case behaviour.
By changing the iteration order, or salting the hash, you can avoid this.
Of course chained hash tables don't suffer from this particular problem.
My quick thought is that hash tables ought keep an internal salt hidden away. This seems good to avoid 'attacks' as well as speeding up merging etc. The only downside I can think of is that the creation of the table needs to fetch a random salt that might not be quick, although that can alleviated by allowing it to be set externally in the table creation so people who don't care can set it to 0 or whatever. What am I missing?
Re: Lessons from Hash Table Merging
#3There's a typo with 'ULL' string suffixes in the hexadecimal numbers in the first code example.
Re: Lessons from Hash Table Merging
#4In Rust, don't do this, it's more work and it'll tend to be slower, often much slower.
HashMap implements Extend, so just h0.extend(h1) and you're done, the people who made your HashMap type are much better equipped to optimize this common operation.
In a new enough C++ in theory you might find the same functionality supported, but Quality of Implementation tends to be pretty frightful.
Re: Lessons from Hash Table Merging
#5In Rust, don't do this, it's more work and it'll tend to be slower, often much slower. HashMap implements Extend, so just h0.extend(h1) and you're done, the people who made your HashMap type are much better equipped to optimize this common operation. In a new enough C++ in theory you might find the same functionality supported, but Quality of Implementation tends to be pretty frightful.
> HashMap implements Extend, so just h0.extend(h1) and you're done, the people who made your HashMap type are much better equipped to optimize this common operation.
Are you sure? I'm not very used to reading Rust stdlib, but this seems to be the implementation of the default HashMap extend [1]. It just calls self.base.extend. self.base seems to be hashbrown::hash_map, and this is the source for it's extend [2]. In other words, does exactly the same thing, just iterates through hash map and inserts it.
Maybe I'm misreading something going through the online docs, or Rust does the "random seed" thing that abseil does, but just blinding assuming something doesn't happen "because Rust" is a bit silly.
[1]: https://doc.rust-lang.org/src/std/collections/hash/map.rs.ht...
[2]: https://docs.rs/hashbrown/latest/src/hashbrown/map.rs.html#4...
Re: Lessons from Hash Table Merging
#6There's a typo with 'ULL' string suffixes in the hexadecimal numbers in the first code example.
No, 0xd6e8feb86659fd93ULL is a valid unsigned long long number. With stdint.h you'd get portable suffix macros, which would help on non-Windows, but they do look worse.
Re: Lessons from Hash Table Merging
#7In Rust, don't do this, it's more work and it'll tend to be slower, often much slower. HashMap implements Extend, so just h0.extend(h1) and you're done, the people who made your HashMap type are much better equipped to optimize this common operation. In a new enough C++ in theory you might find the same functionality supported, but Quality of Implementation tends to be pretty frightful.
> HashMap implements Extend, so just h0.extend(h1) and you're done, the people who made your HashMap type are much better equipped to optimize this common operation. Are you sure? I'm not very used to reading Rust stdlib, but this seems to be the implementation of the default HashMap extend [1]. It just calls self.base.extend. self.base seems to be hashbrown::hash_map, and this is the source for it's extend [2]. In o…
Yes, HashMap will by default be randomly seeded in Rust, but also the code you linked intelligently reserves capacity. If h0 is empty, it reserves enough space for all of h1, and if it isn't then it reserves enough extra space for half of h1, which turns out to be a good compromise.
Note that the worst case is we ate a single unneeded growth, while the best case is that we avoided N - 1 grows where N may be quite large.
Re: Lessons from Hash Table Merging
#8Kudos, neat digging and writeup that makes us think :) If you merge linear probed tables by iterating in sorted hash order then you are matching the storage order and can congest particular parts of the table and cause the linear probing worse case behaviour. By changing the iteration order, or salting the hash, you can avoid this. Of course chained hash tables don't suffer from this particular problem. My quick thou…
Having a per-table key for the hash function is what siphash authors propose and what many do to combat dos attacks right? For example Rust's default HashMap.
The keys are hidden/secret to the system external to the application.
Re: Lessons from Hash Table Merging
#9Maybe this would be a suitable application for "Fibonacci hashing" [0][1], which is a trick to assign a hash table bucket from a hash value. Instead of just taking the modulo with the hash table size, it first multiplies the hash with a constant value 2^64/phi where phi is the golden ratio, and then takes the modulo.
There may be better constants than 2^64/phi, perhaps some large prime number with roughly equal number of one and zero bits could also work.
This will prevent bucket collisions on hash table resizing that may lead to "accidentally quadratic" behavior [2], while not requiring rehashing with a different salt.
I didn't do detailed analysis on whether it helps on hash table merging too, but I think it would.
[0] https://probablydance.com/2018/06/16/fibonacci-hashing-the-o... [1] https://news.ycombinator.com/item?id=43677122 [2] https://accidentallyquadratic.tumblr.com/post/153545455987/r...
Re: Lessons from Hash Table Merging
#10> I evaluated the following hash table libraries, all based on linear probing.
> Abseil
> Rust standard
> hashbrown
These hash tables are not based on plain linear probing, they use something that's essentially quadratic probing done in chunks. Not sure about the others but they might be doing something similar.