As the author of std::hive/plf::hive, I recently learned about boost::hub through a friend, ran my own benchmarks, and engaged in discussions with the author, Joaquin. While we have some disagreements, we generally agree on several points and have learned from each other. Please note that the following assumptions apply only to the current implementations of plf::hive and boost::hub, and future implementations may differ.
- Hub is generally faster overall for smaller types, while hive typically performs better for very large types.
- Insertions are generally faster with hub, except for large types.
- Erasures are faster with hub.
- Results vary slightly by compiler, but in tests measuring the impact of insertions and erasures on iteration using 48-byte structs, hive outperforms hub unless the churn ratio is high.
- The get_iterator() function performs worse when maximum block capacities are smaller, as more blocks need to be checked to locate the pointer, making hub significantly slower than hive.
- Sorting is faster with hub, except for large numbers of large types.
- According to Joaquin's benchmarks, hive seems significantly faster than hub for 32-bit executables, though I haven't tested this myself.
Moreover, visitation is a technique applicable to any semi-contiguous container such as deques, unrolled lists, etc. It combines iteration with pre-fetching, allowing the container to know when the next block begins. However, this technique has limitations and is best used when performing the same operation on a range of elements, but does not integrate well with standard library routines.
From my benchmarks across clang, gcc, and msvc, isolated benchmarks of insert, erase, and iteration are insufficient to measure long-term performance of a hive-like container due to the accumulation of erasures and reserved blocks. Therefore, I suggest conducting your own benchmarks tailored to your use case, as results may vary significantly on different processors.
Blogger's Review: This comparative analysis provides crucial insights for developers in selecting the appropriate container. It is advisable to conduct benchmarks in real application scenarios to ensure the optimal implementation is chosen. The performance may vary significantly based on data types and usage patterns.