Some notes for the big update
martinus opened this issue · comments
- don't split up benchmark results by hash
- Maybe split up into open address hashing and chained hashing y or node based
- Split up into 2 categories: open address hashing (boost maps, std::unordered_map), and all of them.
- Having a filter for the results would be nice
- Disable zoom? at least make a wider view
- Add one summary page with the geomean of all find & insert benchmarks (except the ctor benchmarks)
- Add a conclusio page:
- Use a reasonable hash that spreads entrophy in upper bits to lower bits. std::hash or boost::hash's identity was and is a bad idea. Doesn't need to withstand randomness tests, mumx seems to be good enough)
- Use a pool allocator (boost or PoolAllocator). It's faster and uses much less RAM.
- Create a sortable table: X axis benchmark, y axis map & hash. With one entry that's the GEOMEAN.
most test case is only one hash map.
pls add a new benchmark like this (many small hash maps)
template<class hash_type>
void multi_small_ife(const std::string& hash_name, const std::vector<keyType>& vList)
{
#if KEY_INT
size_t sum = 0;
const auto hash_size = vList.size() / 1003 + 10;
const auto ts1 = getus();
auto mh = new hash_type[hash_size];
for (const auto& v : vList) {
auto hash_id = ((uint32_t)v) % hash_size;
sum += mh[hash_id].emplace(v, 0).second;
}
for (const auto& v : vList) {
auto hash_id = ((uint32_t)v) % hash_size;
sum += mh[hash_id].count(v);
}
for (const auto& v : vList) {
auto hash_id = ((uint32_t)v) % hash_size;
sum += mh[hash_id].erase(v + v % 2);
}
delete []mh;
#endif
}
@ktprime I don't see what that benchmarks adds that is not already covered by the other benchmarks?
the code is copyed from my bench
https://github.com/ktprime/emhash/blob/master/bench/ebench.cpp
I find bench code, key and value(integer case) is alway same type (int-> int, size_t-> size, uint64_t->uint64_t)
can u add or modify some case with <key, value> pairs like <uint64_t, int32_t>?