What bucket_count value should I use if I know the intended number of map keys?

Question

I'm creating an std::unordered_map which I will immediately proceed to populate with n key-value pairs - and I know n. After that no more elements will be added - I will only be performing lookups.

What, therefore, should I pass as bucket_count to the constructor?

Notes:

I know it's not terribly critical and I could simply not specify anything and it will work.
This is related to, but not a dupe of, What should I pass to unordered_map's bucket count argument if I just want to specify a hash function?)
If it helps your answer, you may assume I want to have a load factor between f_1 and f_2 (known in advance).
I'm using the default hash function, and I don't know what the input is like, but it's unlikely to be adversarial to the hashing..

It very much depends on what you will do with this map afterwards. Will you add other elements, or just read it ? Do you seek speed or space efficiency ? How much collisions does you hash function yield on your set ? It is reasonable to think about the load factor rather than the actual number of buckets. — Jean-Bernard Jansen, Dec 16 '16 at 11:13
@Jean-BernardJansen: See edit. Also, I'd like some sort of a reasonable default - the same way we now have a reasonable default without knowing n. Adding that information, and applying the same considerations, should yield some number... — einpoklum, Dec 16 '16 at 11:21

Martin Bonner supports Monica · Accepted Answer · 2016-12-16T11:25:18.010

According to n4296 in 23.5.4.2 [unord.map.cnstr] (this is the final draft of C++14) by default the max_load_factor for an unordered_map is 1.0, so you could just set the bucket_count to n.

There is obviously a space-time trade-off between increasing the bucket count for improved speed and decreasing it (and raising the max load factor) for improved space.

I would either not worry about it, or if it is a large map, set the bucket count to n. Then you can worry about optimizing when profiling shows you have a problem.

If you know the range of load factors you want, then you just set the bucket count to std::ceil(n/(std::max(f_1,f_2)), (and set the load factor before you fill the map).

score 0 · Answer 2 · answered Dec 16 '16 at 11:26

Given the fact you have a range for your load factor, the only missing information is the collision rate. You can simply use nb_buckets = n / f_2 and you will be sure to get a load factor less than or equal to f_2. Ensuring correctness about f_1 requires data about the collision rate.

What bucket_count value should I use if I know the intended number of map keys?

2 Answers2