W dniu 25 marca 2010 23:25 użytkownik Joshua A. Andler <scislac@...400...> napisał:
I just ran across the following and figured I would pass the link on since it is a comparison of various hash map implementations and their performance.
http://incise.org/hash-table-benchmarks.html
Are we using unordered w/ integers? If so, we may want to explore other options.
Oh, and someone in the notes also pointed out:
https://edge.launchpad.net/libmct
Cheers, Josh
These benchmarks indicate that the GCC implementation fares well enough that replacing it with anything else is hardly worth the effort. We'll exhaust all available memory in the renderer and SP tree long before hash table performance begins to be significant. We currently exhaust the 32-bit address space on SVGs as small as several megabytes, see
https://bugs.launchpad.net/inkscape/+bug/168738
Such SVGs cannot have more than 2-3 million nodes. The node tool uses 1 pointer hash entry per unselected node and 2 per selected node.
The order of preference as of now is: TR1 (GCC) > Boost > __gnu_cxx::hash_map. I can change it to TR1 > __gnu_cxx > Boost, since it seems that the Boost implementation is worst.
Regards, Krzysztof