On Wed, 24 Jan 2024 00:26:09 GMT, Joshua Cao <d...@openjdk.org> wrote:
> This change mirrors what we did for ConcurrentHashMap in > https://github.com/openjdk/jdk/pull/17116. When we add all entries from one > map to anther, we should resize that map to the size of the sum of both maps. > > I used the command below to run the benchmarks. I set a high heap to reduce > garbage collection noise. > > java -Xms25G -jar benchmarks.jar -p size=100000 -p addSize=100000 -gc true > org.openjdk.bench.java.util.HashMapBench > > > Before change > > > Benchmark (addSize) (mapType) (size) Mode Cnt Score > Error Units > HashMapBench.putAll 100000 HASH_MAP 100000 avgt 4 22.927 ± > 3.170 ms/op > HashMapBench.putAll 100000 LINKED_HASH_MAP 100000 avgt 4 25.198 ± > 2.189 ms/op > > > After change > > > Benchmark (addSize) (mapType) (size) Mode Cnt Score > Error Units > HashMapBench.putAll 100000 HASH_MAP 100000 avgt 4 16.780 ± > 0.526 ms/op > HashMapBench.putAll 100000 LINKED_HASH_MAP 100000 avgt 4 19.721 ± > 0.349 ms/op > > > We see about average time improvements of 26% in HashMap and 20% in > LinkedHashMap. Then we might need some statistics on how often `putAll` replaces existing mappings, ranging from none at all to completely. For example, `Collectors.toMap` would never replace existing mappings, even though it doesn't use `putAll` (it can probably call putAll and throw exception if the new size isn't 2 old sizes added up) The current allocation assumes putting replaces all existing mappings, which I don't think is quite applicable. ------------- PR Comment: https://git.openjdk.org/jdk/pull/17544#issuecomment-1908819980
