Performance of array of non-nullable value objects

[David Witten]

First I'd like to thank everyone working on valhalla, it is very exciting!

I was playing around with different hash Map implementations using Build
20-valhalla+20-75
<https://download.java.net/java/early_access/valhalla/20/openjdk-20-valhalla+20-75_linux-x64_bin.tar.gz>.
I was surprised that get() performance when bins were arrays of primitive
classes containing (int hash, K key, V value) performed worse than the
normal non-primitive linked-list bins of (int hash, K key, V value, Node
next) even when bins were rather large.

Is this expected to be the case for this old early release?
Is this expected to be the case when non-nullable, not-atomicly-updated
value objects are officially released?
Is there any guess about when another early release will be available?

The initial size and load factor of the Maps guaranteed that the bins were
largish.  The random seed was the same for array-bin vs list-bin so that
the bins sizes are always same for both.  Here is a typical histogram for
bin sizes (again same for list vs array bins):

histogram: max: 16, [589, *2436*, 4805, 6396, 6342, 5173, 3413, 1889, 1017,
437, 187, 55, 21, 6, 1, 0, 1]
So, for example, there were *2436* bins of size 1.

Here are some results (I started with the jmh tests in
https://hg.openjdk.org/valhalla/valhalla/rev/71ba79398dd9 ) but the Map
implementations and random seed approach are different.

*Benchmark                              (mapType)  (size)  Mode  Cnt
Score     Error  Units*
GetX.getHit  mapprotos.ArrayBinHashMapJustPutGet  131072  avgt   10
 4784.596 ± 649.247  us/op
GetX.getHit          mapprotos.HashMapJustPutGet  131072  avgt   10
 3053.680 ± 427.430  us/op
GetX.getMix  mapprotos.ArrayBinHashMapJustPutGet  131072  avgt   10
 4991.512 ± 292.345  us/op
GetX.getMix          mapprotos.HashMapJustPutGet  131072  avgt   10
 4291.636 ± 543.059  us/op

If this sort of behavior is not expected, and someone would like to look
into it, I can clean up and supply my test.


-thanks, David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/valhalla-dev/attachments/20240213/c2ec8ee6/attachment.htm>