Table of Contents

I’ve spent a long time developing my robin_hood::unordered_map, and after claiming that it is now the fastest hashmap I understandably got quite a few skeptic comments. Some of the comments were quite right, and my benchmarks were not as unbiased as they could be, I did not test as many unordered maps as I should have, my compiler options were not choosen well, and so on.

That’s why I have now spent considerable time to create a highly improved benchmarks, where I have tried to remedy all most of the critique that I got. The results are not as flattering to my robin_hood::unordered_map, but I am still very pleased with the results.

What is actually Benchmarked?

This benchmark has evalued 20 different unordered_map implementations, each with 5 different hashing implementations. So there are a total of 20*5 = 100 hashmap variants to benchmark. Each of this 100 hashmaps was evaluated in 10 different benchmarks, so in total 1000 benchmark evaluations. I ran each benchmark 9 times and show the median, to filter out any outliers. So in total I ran 9000 benchmarks, which took about 6 days on my Intel i7-8700 at 3200 MHz. To get highly accurate results, I’ve isolated a core to only benchmarking, and disabled all frequency scaling.

Hashmaps

  • Google’s Abseil’s abseil::flat_hash_map, abseil::node_hash_map. They are brand new, and have just recently pushed the boundary on what’s possible to achieve for unordered_maps. It uses several interesting optimizations, described in CppCon 2017: Matt Kulukundis “Designing a Fast, Efficient, Cache-friendly Hash Table, Step by Step.
  • boost multiindex: boost::multi_index::hashed_unique. Boost.MultiIndex is a versatile container that is highly configurable, it’s main features is not speed but it’s versatility. It is not a straight forward std::unordered_map replacement, the implementation for the wrapper was thankfully provided by joaquintides.
  • Boost’s unordered map boost::unordered_map is very similar to std::unordered_map, just boosts (older) version before std::unordered_map was a thing. I’ve tested with boost version 1.65.1.
  • EASTL has eastl::hash_map. The Electronic Arts Standard Template Library, an STL implementation with emphasis on high performance. It seems to be a bit dated though.
  • Facebook’s folly: folly::F14ValueMap and folly::F14NodeMap. C++14 conform and high performance in mind. The maps are described in the F14 Hash Table document.
  • greg7mdp’s parallel hashmap: phmap::flat_hash_map and phmap::node_hash_map are closely based on Abseil’s map, but simpler to integrate since they are header only. phmap::parallel_flat_hash_map and phmap::parallel_node_hash_map use a novel improvement that makes the maps a tad slower but usable in parallel. Also, peak memory requirements are a bit lower. Read more in “The Parallel Hashmap”.
  • greg7mdp’s sparsepp: spp::sparse_hash_map tuned to be memory efficient.
  • ktprime’s HashMap: A rather unknown implementation emilib1::HashMap by /u/huangyuanbing. It might not be as stable and well tested as other implementations here, but the numbers look very promising.
  • martinus’s robin-hood-hashing: A single-file header-only implementation that contains robin_hood::unordered_flat_map and robin_hood::unordered_node_map. I am the author of the maps, so I might not be perfectly unbiased… The numbers won’t lie though, and I try to be as objective as possible.
  • Malte Skarupke’s Bytell Map After first claiming I Wrote The Fastest Hashtable and later A new fast hash table in response to Google’s new fast hash table, his maps ska::flat_hash_map and ska::bytell_hash_map are obvious choices for this benchmark.
  • std::unordered_map Of course, the standard implementation of std::unordered_map has to be included has well Since I am using g++ 8.2, this uses the libstdc++ implementation.
  • tessil’s maps: Tessil has done lots and lots of work on hashmaps, in all kinds of flavours. Here I am benchmarking tsl::hopscotch_map, tsl::robin_map, and tsl::sparse_map. They are all available on github.

Hashes

Some hashmap implementations come with their own hashing methods, each with different properties. In my benchmarks I have used either integral types like int or uint64_t, and std::string as the keys.

  • Abseil’s Hash absl:Hash: An extremely fast hash, that works very well in all situations I have tested.
  • FNV1a A very simple hash that is used by Microsoft in Visual Studio 2017. Interestingly, they even use this byte-wise hash for integral types. My benchmark has its own implementation, but in my experiments it has produced the same assembler code as the original Microsoft variant.
  • Folly’s Hash folly::hasher: Unfortunately I could not find any documentation. It seems to be well optimized and uses native crc instruction if available. Unfortunately the result is only a 32bit hash which can work badly for some hashmap variants.
  • libstdc++-v3 simply casts integral types to size_t and uses this as a hash function. It is obviously the fastest hash, but many hashmap implementations rely on a somewhat good avalanching hash quality so this seems to be a rather bad choice.
  • martinus’s robin-hood-hashing robin_hood::hash is based on abseil’s hash for integral types, with minor modifications.

How is benchmarked?

Build

  • I’ve used g++ 8.2.0 with -O3 -march=native: 
    g++-8 (Ubuntu 8.2.0-1ubuntu2~18.04) 8.2.0
  • CMake build is done with Release mode
  • and I’ve set FOLLY_CXX_FLAGS to -march=native.
  • For the ktprime map benchmarks I had to add -fno-strict-aliasing.

System Configuration

  • All benchmarks were run on an Linux. uname -a output: 
    Linux dualit 4.15.0-47-generic #50-Ubuntu SMP Wed Mar 13 10:44:52 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux
  • Processor Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz, locked to 3200 MHz.
  • Isolated a core with it’s hyperthreading companion by editing /etc/default/grub and changing GRUB_CMDLINE_LINUX_DEFAULT so it looks like this: 
    GRUB_CMDLINE_LINUX_DEFAULT="quiet splash isolcpus=5,11 rcu_nocbs=5,11"
  • Turbo boost and frequency scaling were disabled with the python tool perf with the command 
    sudo python3 -m perf system tune

    This sets cores 5 and 11 to 3200 MHz, sets scaling governor to performance, disables Turbo Boost, sets irqbalances service to inactive, IRQ affinity to all CPUs except 5 and 11.

  • Each benchmarks is run in a separately started process.
  • Isolated cores are used with taskset -c 5,11
  • To get rid of any potential outliers and to average effects of ASLR, all benchmarks were run 9 times and I show only the median result.

Benchmarks

Enough talk, onwards to the benchmarks!