RFR: 8322149: ConcurrentHashMap smarter presizing for copy constructor and putAll [v3]

[jmehrens]

On Wed, 17 Jan 2024 21:16:02 GMT, Joshua Cao <duke at openjdk.org> wrote:

>> ConcurrentHashMap's copy constructor calls `putAll()` -> `tryPresize()` -> `transfer()`. When coming from the copy constructor, the Map is empty, so there is nothing to transfer. But `transfer()` will still copy all the empty nodes from the old table to the new table.
>> 
>> This patch avoids this work by only calling `tryPresize()` if the table is already initialized. If `table` is null, the initialization is deferred to `putVal()`, which calls `initTable()`.
>> 
>> ---
>> 
>> ### JMH results for testCopyConstructor
>> 
>> before patch:
>> 
>> 
>> Result "org.openjdk.bench.java.util.concurrent.Maps.testCopyConstructor":
>>   937395.686 ±(99.9%) 99074.324 ns/op [Average]
>>   (min, avg, max) = (825732.550, 937395.686, 1072024.041), stdev = 92674.184
>>   CI (99.9%): [838321.362, 1036470.010] (assumes normal distribution)
>> 
>> 
>> after patch:
>> 
>> 
>> Result "org.openjdk.bench.java.util.concurrent.Maps.testCopyConstructor":
>>   620871.469 ±(99.9%) 59195.406 ns/op [Average]
>>   (min, avg, max) = (545304.633, 620871.469, 689013.573), stdev = 55371.419
>>   CI (99.9%): [561676.063, 680066.875] (assumes normal distribution)
>> 
>> 
>> Average time is decreased by about 33%.
>> 
>> ### JMH results for testPutAll (size = 10000)
>> 
>> before patch:
>> 
>> 
>> Result "org.openjdk.bench.java.util.concurrent.Maps.testConcurrentHashMapPutAll":
>>   4315291.542 ±(99.9%) 336034.190 ns/op [Average]
>>   (min, avg, max) = (3974688.194, 4315291.542, 4871772.209), stdev = 314326.589
>>   CI (99.9%): [3979257.352, 4651325.731] (assumes normal distribution)
>> 
>> 
>> after patch:
>> 
>> 
>> Result "org.openjdk.bench.java.util.concurrent.Maps.testConcurrentHashMapPutAll":
>>   3006955.723 ±(99.9%) 271757.969 ns/op [Average]
>>   (min, avg, max) = (2801264.198, 3006955.723, 3553084.135), stdev = 254202.573
>>   CI (99.9%): [2735197.754, 3278713.692] (assumes normal distribution)
>> 
>> 
>> Average time is decreased about 30%.
>
> Joshua Cao has updated the pull request incrementally with one additional commit since the last revision:
> 
>   putAll presize based on sum on both map sizes

src/java.base/share/classes/java/util/concurrent/ConcurrentHashMap.java line 1088:

> 1086:     public void putAll(Map<? extends K, ? extends V> m) {
> 1087:         if (table != null) {
> 1088:             tryPresize(size() + m.size());

Is overflow not an issue here because calling tryPresize with a negative value will invoke tableSizeFor?

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/17116#discussion_r1456919954