didn't measure difference on stock exchange HFT data stream about sparsepp HOT 5 CLOSED

Hi @steven-varga. To help me understand exactly what your use case is, can you post the declaration of the sparse_hash_map that you used?
I have the feeling that the memory used by the hash table to index the 6000 symbols is negligible in comparison of the data stored (the 6e7 count/day), which doesn't depend on the hash table implementation. Thanks!

from sparsepp.

Yes indeed: it is about speed as opposed to footprint.
spp::sparse_hash_map<std::string, uint64_t> map1; or spp::sparse_hash_map<uint64_t, uint64_t> map2;
Then I used auto it = map.find( key ); and inserted new ones upon failure.
The profiling showed the code path spends 50% on hash lookups, I tried a quick drop in, and didn't notice difference.

from sparsepp.

Ah, sparsepp is mostly about footprint, especially when inserting entries and resizing occurs (high water mark). Should be a little bit faster, though!

from sparsepp.

update: the attached pictures were identical, corrected mistake. The result is same runtime.
Profiling and runtime attached, done on: gcc (Ubuntu 5.4.0-6ubuntu1~16.04.5) 5.4.0 20160609: Intel(R) Core(TM) i7-3520M CPU @ 2.90GHz, runtime is near same, slight 1s difference when using std::map
hope it helps!

from sparsepp.

Hum, I don't know what to say here. The implementation of sparsepp require more copies (or move) of the hash table objects than a regular hash table, so for complex objects which are expensive to copy or move, sparsepp may not be the best alternative.

from sparsepp.