robin-hood-hashing VS parallel-hashmap

From your other comments, it seems like your knowledge of hash tables might be limited to closed-addressing/separate-chaining hash tables. The current frontrunners in high-performance, memory-efficient hash table design all use some form of open addressing, largely to avoid pointer chasing and limit cache misses. In this regard, you want to check our SSE-powered hash tables (such as Abseil, Boost, and Folly/F14), Robin Hood hash tables (such as Martinus and Tessil), or Skarupke (I've recently had a lot of success with a similar design that I will publish here soon and is destined to replace my own Robin Hood hash tables). Also check out existing research/benchmarks here and here. But we a little bit wary of any benchmarks you look at or perform because there are a lot of factors that influence the result (e.g. benchmarking hash tables at a maximum load factor of 0.5 will produce wildly different result to benchmarking them at a load factor of 0.95, just as benchmarking them with integer keys-value pairs will produce different results to benchmarking them with 256-byte key-value pairs). And you need to familiarize yourself with open addressing and different probing strategies (e.g. linear, quadratic) first.

Also, FYI there is robin_hood::unordered_{map,set} which has very high performance, and is header-only and standalone.

std::unordered_map is notoriously slow, several times slower than a "proper" hashmap implementation like Google's absl or Martin's robin-hood-hashing [1]. That said, std::sort is not the fastest sort implementation, either. It is hard to say which will win.

[1]: https://github.com/martinus/robin-hood-hashing

The main advantage of the latest version is that it reduces build time by about 53% (GCC 12.1), based on the comprehensive test suit found in unit_tests.c. This improvement is significant because compile time was previously a drawback of this library, with maps and sets—in particular—compiling slower than their C++ template-based counterparts. I achieved it by refactoring the library to do less work inside API macros and, in particular, use fewer _Generic statements, which seem to be a compile-speed bottleneck. A nice side effect of the refactor is that the library can now more easily be extended with the planned dynamic strings and ordered maps and sets. The other major improvement concerns the performance of maps and sets. Here are some interactive benchmarks[1] comparing CC’s maps to two popular implementations of Robin Hood hash maps in C++ (as well as std::unordered_map as a baseline). They show that CC maps perform roughly on par with those implementations.

For anyone in a situation where a set/map (or unordered versions) is in a hot part of the code, I'd also highly recommend Robin Hood: https://github.com/martinus/robin-hood-hashing

It made a huge difference in one of the programs I was running.

Oh my bad. Still thought -- your name.. it looks very familiar to me. Are you the robin_hood hashing guy perhaps? Yes you are! My bad -- https://github.com/martinus/robin-hood-hashing.

Got a bit better C++ version here which uses a couple libraries instead of std:: stuff - https://gist.github.com/jcelerier/74dfd473bccec8f1bd5d78be5a... ; boost, fmt and https://github.com/martinus/robin-hood-hashing

    $ g++ -I robin-hood-hashing/src/include -O2 -flto -std=c++20 -fno-exceptions -fno-unwind-tables -fno-asynchronous-unwind-tables -lfmt

The implementation is mostly inspired by this comment and lessons learned from my older robin-hood-hashing hashmap.

My default they are not thread safe, i.e., they offer the same thread safety as std::map or any stdlib type; however, the map can optionally be made thread safe and is apparently optimized for this usage. Details at: https://github.com/greg7mdp/parallel-hashmap?tab=readme-ov-f....

This looks pretty good in terms of its tradeoffs and tricks used. Makes good use of the Metadata, as in https://greg7mdp.github.io/parallel-hashmap/

Standard library maps/unordered_maps are themselves notoriously slow anyway. A sparse_hash_map from abseil or parallel-hashmaps[1] would be better.

[1] https://github.com/greg7mdp/parallel-hashmap

Cool! Looking forward to you trying my phmap - and please let me know if you have any question.

I do this as well in my phmap and gtl implementations. It makes the tables look worse in benchmarks like the above, but prevents really bad surprises occasionally.

Thanks a lot for the great benchmark, Martin. Glad you used different hash functions, because I do sacrifice some speed to make sure that the performance of my hash maps doesn't degrade drastically with poor hash functions. Happy to see that my phmap and gtl (the C++20 version) performed well.

I use https://github.com/greg7mdp/parallel-hashmap when I need better performance of maps and sets.

Then they should really try https://github.com/greg7mdp/parallel-hashmap, the current state of the art.

std::unordered_map is notorious for being slow. Use a better implementation (I like the flat naps from here, which are the same as abseil’s). The question that needs to be asked too is if you need to use a map.