Intel® oneAPI Threading Building Blocks
Ask questions and share information about adding parallelism to your applications when using this threading library.
2465 Discussions

concurrent_unordered_map: good hash functionsj for lists of 32-bit integers, character strings, etc


I am disappointed with the lack of documentation regarding the hashing for concurrent_unordered_map. In particular, as a drop in replacement for boost's concurrent_unordered_map, they have completely different hash behavior.

In my application, I hash sorted vectors of 32-bit integers (std::size_t or unsigned int), With Boost I was using the 'PJW Hash Function'
From page 436 of the Aho Compiler Book (The one with the dragon on the cover)
Also seen here:

The boost::unordered_map uses chained addressing and implements the canonical buckets with linked lists, and uses a prime number for the bucket sizes. You can see the large amount of explanation on it here:

the concurrent_unordered_map uses a skip-order list. Essentially the data structure is a linked list of all your data, and the "buckets" are pointers into specific entries of this list. To be honest, even after reading this paper late last night, I'm still a bit confused as to the particular invariants and implementations, but, the biggest important difference, is that the paper seems to use a table size of size 2^k.
Consider (by example), taking any key (unsigned int) and reducing it mod 2^k. For example, if k = 1, then only the least significant bit of the key matters for determining a bucket. i.e. the lowest k significant bits determine the bucket. So a good hash here should enforce randomness on those k bits.

I'll be posting more to this soon with more explanation as I have it. I hope that those particularly well informed, could post more on this matter.

One question I have, is, if I know in advance the total size of this hash table, how should I choose the buckets as to minimize space and search time?
0 Kudos
1 Reply
Valued Contributor III
With concurrent_hash_map you need good randomness in the low-order bits, so you could try if that also helps with concurrent_unordered_map.
0 Kudos