Our great sponsors
-
zig
General-purpose programming language and toolchain for maintaining robust, optimal, and reusable software.
-
WorkOS
The modern identity platform for B2B SaaS. The APIs are flexible and easy-to-use, supporting authentication, user identity, and complex enterprise features like SSO and SCIM provisioning.
-
tigerbeetle
The distributed financial transactions database designed for mission critical safety and performance.
-
-
robin-hood-hashing
Discontinued Fast & memory efficient hashtable based on robin hood hashing for C++11/14/17/20
-
InfluxDB
Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.
Thanks for the details on your benchmarks. I would like sometime to extend BLP to a more generic setting; as I said I think any trick used with RH would also work with BLP. I just used an integer set because that's all I needed for my use case and it was easy to implement several different approaches for benchmarking. As you note, it favors use cases where the hash function is cheap (or invertible) and elements are cheap to move around.
About your question on load factors: no, the benchmarks are measuring exactly what they claim to be. The hash table constructor divides max data size by load factor to get the table size (https://github.com/senderista/hashtable-benchmarks/blob/mast...), and the benchmark code instantiates each hash table for exactly the measured data set size and load factor (https://github.com/senderista/hashtable-benchmarks/blob/mast...).
I can't explain the peaks around 1M in many of the plots; I didn't investigate them at the time and I don't have time now. It could be a JVM artifact, but I did try to use JMH "best practices", and there's no dynamic memory allocation or GC happening during the benchmark at all. It would be interesting to port these tables to Rust and repeat the measurements with Criterion. For more informative graphs I might try a log-linear approach: divide the intervals between the logarithmically spaced data sizes into a fixed number of subintervals (say 4).
Actually it seems according to the issue that TigerBeetle (one of the bigger zig projects out there) noticed this issue [1]. It's also on their issue tracker [2].
[1] https://github.com/ziglang/zig/issues/17851
My impression from the article is that Zig provides several different hashtables and not all of them are broken in this way.
This reminds me of Aria's comment in her Rust tutorial https://rust-unofficial.github.io/too-many-lists/ about failing to kill LinkedList. One philosophy (and the one Rust chose) for a stdlib is that this is only where things should live when they're so commonly needed that essentially everybody needs them either directly or to talk about. So, HashTable is needed by so much otherwise unrelated software that qualifies, BloomFilter, while it's real useful for some people, not so much. Aria cleaned out Rust's set of standard library containers before Rust 1.0, trying to keep only those most people would need. LinkedList isn't a good general purpose data structure, but, it was too popular and Aria was not able to remove it.
Having multiple hash tables feels like a win (they're optimized for different purposes) but may cost too much in terms of the necessary testing to ensure they all hit the quality you want.
In my example the table stores the hash codes themselves instead of the keys (because the hash function is invertible)
Oh, I see, right. If determining the home bucket is trivial, then the back-shifting method is great. The issue is just that it’s not as much of a general-purpose solution as it may initially seem.
“With a different algorithm (Robin Hood or bidirectional linear probing), the load factor can be kept well over 90% with good performance, as the benchmarks in the same repo demonstrate.”
I’ve seen the 90% claim made several times in literature on Robin Hood hash tables. In my experience, the claim is a bit exaggerated, although I suppose it depends on what our idea of “good performance” is. See these benchmarks, which again go up to a maximum load factor of 0.95 (Although boost and Absl forcibly grow/rehash at 0.85-0.9):
https://strong-starlight-4ea0ed.netlify.app/
Tsl, Martinus, and CC are all Robin Hood tables (https://github.com/Tessil/robin-map, https://github.com/martinus/robin-hood-hashing, and https://github.com/JacksonAllan/CC, respectively). Absl and Boost are the well-known SIMD-based hash tables. Khash (https://github.com/attractivechaos/klib/blob/master/khash.h) is, I think, an ordinary open-addressing table using quadratic probing. Fastmap is a new, yet-to-be-published design that is fundamentally similar to bytell (https://www.youtube.com/watch?v=M2fKMP47slQ) but also incorporates some aspects of the aforementioned SIMD maps (it caches a 4-bit fragment of the hash code to avoid most key comparisons).
As you can see, all the Robin Hood maps spike upwards dramatically as the load factor gets high, becoming as much as 5-6 times slower at 0.95 vs 0.5 in one of the benchmarks (uint64_t key, 256-bit struct value: Total time to erase 1000 existing elements with N elements in map). Only the SIMD maps (with Boost being the better performer) and Fastmap appear mostly immune to load factor in all benchmarks, although the SIMD maps do - I believe - use tombstones for deletion.
I’ve only read briefly about bi-directional linear probing – never experimented with it.
In my example the table stores the hash codes themselves instead of the keys (because the hash function is invertible)
Oh, I see, right. If determining the home bucket is trivial, then the back-shifting method is great. The issue is just that it’s not as much of a general-purpose solution as it may initially seem.
“With a different algorithm (Robin Hood or bidirectional linear probing), the load factor can be kept well over 90% with good performance, as the benchmarks in the same repo demonstrate.”
I’ve seen the 90% claim made several times in literature on Robin Hood hash tables. In my experience, the claim is a bit exaggerated, although I suppose it depends on what our idea of “good performance” is. See these benchmarks, which again go up to a maximum load factor of 0.95 (Although boost and Absl forcibly grow/rehash at 0.85-0.9):
https://strong-starlight-4ea0ed.netlify.app/
Tsl, Martinus, and CC are all Robin Hood tables (https://github.com/Tessil/robin-map, https://github.com/martinus/robin-hood-hashing, and https://github.com/JacksonAllan/CC, respectively). Absl and Boost are the well-known SIMD-based hash tables. Khash (https://github.com/attractivechaos/klib/blob/master/khash.h) is, I think, an ordinary open-addressing table using quadratic probing. Fastmap is a new, yet-to-be-published design that is fundamentally similar to bytell (https://www.youtube.com/watch?v=M2fKMP47slQ) but also incorporates some aspects of the aforementioned SIMD maps (it caches a 4-bit fragment of the hash code to avoid most key comparisons).
As you can see, all the Robin Hood maps spike upwards dramatically as the load factor gets high, becoming as much as 5-6 times slower at 0.95 vs 0.5 in one of the benchmarks (uint64_t key, 256-bit struct value: Total time to erase 1000 existing elements with N elements in map). Only the SIMD maps (with Boost being the better performer) and Fastmap appear mostly immune to load factor in all benchmarks, although the SIMD maps do - I believe - use tombstones for deletion.
I’ve only read briefly about bi-directional linear probing – never experimented with it.
In my example the table stores the hash codes themselves instead of the keys (because the hash function is invertible)
Oh, I see, right. If determining the home bucket is trivial, then the back-shifting method is great. The issue is just that it’s not as much of a general-purpose solution as it may initially seem.
“With a different algorithm (Robin Hood or bidirectional linear probing), the load factor can be kept well over 90% with good performance, as the benchmarks in the same repo demonstrate.”
I’ve seen the 90% claim made several times in literature on Robin Hood hash tables. In my experience, the claim is a bit exaggerated, although I suppose it depends on what our idea of “good performance” is. See these benchmarks, which again go up to a maximum load factor of 0.95 (Although boost and Absl forcibly grow/rehash at 0.85-0.9):
https://strong-starlight-4ea0ed.netlify.app/
Tsl, Martinus, and CC are all Robin Hood tables (https://github.com/Tessil/robin-map, https://github.com/martinus/robin-hood-hashing, and https://github.com/JacksonAllan/CC, respectively). Absl and Boost are the well-known SIMD-based hash tables. Khash (https://github.com/attractivechaos/klib/blob/master/khash.h) is, I think, an ordinary open-addressing table using quadratic probing. Fastmap is a new, yet-to-be-published design that is fundamentally similar to bytell (https://www.youtube.com/watch?v=M2fKMP47slQ) but also incorporates some aspects of the aforementioned SIMD maps (it caches a 4-bit fragment of the hash code to avoid most key comparisons).
As you can see, all the Robin Hood maps spike upwards dramatically as the load factor gets high, becoming as much as 5-6 times slower at 0.95 vs 0.5 in one of the benchmarks (uint64_t key, 256-bit struct value: Total time to erase 1000 existing elements with N elements in map). Only the SIMD maps (with Boost being the better performer) and Fastmap appear mostly immune to load factor in all benchmarks, although the SIMD maps do - I believe - use tombstones for deletion.
I’ve only read briefly about bi-directional linear probing – never experimented with it.
In my example the table stores the hash codes themselves instead of the keys (because the hash function is invertible)
Oh, I see, right. If determining the home bucket is trivial, then the back-shifting method is great. The issue is just that it’s not as much of a general-purpose solution as it may initially seem.
“With a different algorithm (Robin Hood or bidirectional linear probing), the load factor can be kept well over 90% with good performance, as the benchmarks in the same repo demonstrate.”
I’ve seen the 90% claim made several times in literature on Robin Hood hash tables. In my experience, the claim is a bit exaggerated, although I suppose it depends on what our idea of “good performance” is. See these benchmarks, which again go up to a maximum load factor of 0.95 (Although boost and Absl forcibly grow/rehash at 0.85-0.9):
https://strong-starlight-4ea0ed.netlify.app/
Tsl, Martinus, and CC are all Robin Hood tables (https://github.com/Tessil/robin-map, https://github.com/martinus/robin-hood-hashing, and https://github.com/JacksonAllan/CC, respectively). Absl and Boost are the well-known SIMD-based hash tables. Khash (https://github.com/attractivechaos/klib/blob/master/khash.h) is, I think, an ordinary open-addressing table using quadratic probing. Fastmap is a new, yet-to-be-published design that is fundamentally similar to bytell (https://www.youtube.com/watch?v=M2fKMP47slQ) but also incorporates some aspects of the aforementioned SIMD maps (it caches a 4-bit fragment of the hash code to avoid most key comparisons).
As you can see, all the Robin Hood maps spike upwards dramatically as the load factor gets high, becoming as much as 5-6 times slower at 0.95 vs 0.5 in one of the benchmarks (uint64_t key, 256-bit struct value: Total time to erase 1000 existing elements with N elements in map). Only the SIMD maps (with Boost being the better performer) and Fastmap appear mostly immune to load factor in all benchmarks, although the SIMD maps do - I believe - use tombstones for deletion.
I’ve only read briefly about bi-directional linear probing – never experimented with it.