TripleA Performance with Java 26

[M0ON-MAN]

Test machine game installed on:

• Framework16 Ryzen Ai9 HX370
• 64GB 5600 CL40 RAM
• 8TB Samsung Pro 9100 SSD
• Java 26.0.1
• Win 11 Pro

Game round = 618

TripleA.vmoptions file entries:
-Xms20G
-Xmx20G
-XX:ActiveProcessorCount=12
-XX:+UseZGC
-XX:+AlwaysPreTouch
-XX:SoftRefLRUPolicyMSPerMB=0
-XX:ConcGCThreads=6
-XX:+ZProactive
-Djava.util.concurrent.ForkJoinPool.common.parallelism=8
-Dsun.java2d.opengl=false
-Dsun.java2d.d3d=false
-Dsun.java2d.noddraw=true
-Dsun.java2d.vulkan=true
-Dsun.java2d.vulkan.fbobject=true
-Dfile.encoding=UTF-8

triplea-2026-05-20.1.log

[TheDog-GH]

@M0ON-MAN
TripleA 2.7.15566 comes with Java 21.0.7 and you have custom TripleA.vmoptions file,
I do too, it only has
-Xmx4096M
-Xms4096M
-Xss2048K

.
@beelee1 Yes Im on Windows 11

[M0ON-MAN]

@TheDog-GH I have been playing it for over a year with Java 25.0.2 and now 26.0.1 without any issues and able to take advantage of the additional flags. You can see a am in turn 618! That is a lot of testing on Java 25 and 26 versions need to update java to 26 because 21 is too old and not taking advantage of all the massive enhancements and performance gains that come with 26.

[TheDog-GH]

I will guess your Xss needs to be increased ?

[DanVanAtta]

Thank you for running through the test cases. @TheDog-GH had a good eye that 'Xss' is needed on a few maps. The 20GB heap size is pretty extreme! The issue with heaps that are "too" large is that garbage collection pauses can become very slow. The time to do garbage collection is super linear relative to the size of the heap (meaning, doubling the heap size can quadruple or more the time to do garbage collection). Generally going over 12GB becomes a 'danger' territory, though, you have a beast of a machine and likely did not notice.

Question: did you find that you needed a heap that large?

RE: Java 26

There are three things we really need to be sure line up in order to upgrade java:

• the JRE bundled with installj4 needs to be available for us to pick it up. Install4j is the game installer.
• the dev tools, specifically the various gradle plugins that we use, need to all be compatible.
• library dependencies need to be compatible. AFAIK we still make pretty extensive use of Guava, which tends to not do well on Java version updates. I wouldn't mind seeing us rip out Guava to help avoid that issue, but that is a prime suspect of a library that might not allow us to upgrade cleanly

[M0ON-MAN]

@DanVanAtta When i play with the stock Java 21 my AI turns take 2-3 time more to complete (yes I have timed them with a stop watch) and using the exact same HW. When I switch to Java 25 or 26 the AI turns 2-3 times faster (again timed with stop watch). Java 26 Garbage Collector was totally revamped it is a massive improvement over Java 21 and you can throw up to 40GB for xms/amx and it will still give you about exactly the same timings. I settled for 20GB that was the perfect "sweet spot". Again I am on turn 619! that is a "massive" amount of data to handle and Java 25 and 26 can handle it much more efficiently than Java 21. Again this is thru 6-7 months of me testing, tinkering with Java 25 and 26. I am going to continue pushing it. Recently I started utilizing Claude AI and have been tweaking the Java 26 flags with its help and timing everything and comparing logs and output. Again, my laptop (Framework 16) machine is not what the average user out there has so maybe my experience and performance are an outlier... but it is fun and I will continue to test and share :) Thank you!

PS:
I am in the process of building a Framework Desktop machine with 128GB ram so I will test it a save starting at turn 600 running in a VM so I can dig deeper in all the processes and how Java 26 is interacting and handling the threads and memory utilization. I have to "FORCE" Java to use additional threads and cores using:
-XX:ActiveProcessorCount=12
-XX:ConcGCThreads=6
-Djava.util.concurrent.ForkJoinPool.common.parallelism=8

Without the above flags the game just latches on to 1 or 2 cores only and I have 12 cores. With the flags it uses 4-6 additional cores... need to add additional logic in the games code to utilize more cores/threads.

[DanVanAtta]

I converted this issue to a discussion.

@M0ON-MAN interesting, a few things I'm curious about:

• which garbage collector are you using? A default one, or did you specify one of the alternatives? There are multiple to choose from
• The number of garbage collections and the timing of full and partial collections would be interesting. Claude should be able to help you get those stats. With the 20GB heap, you might never be seeing full GCs at all. Those could be really slow potentially (or not).
• A typical player will likely have anywhere from 4GB of memory to 16GB. It can be really worthwhile to find settings that are optimized for 8GB so that we could adopt those as defaults that would apply to the full user base. At the same time, knowing that approaching 20GB is a sweet spot is very valuable too. Which leads to the question, are the 'best' settings for a nominal 8GB of memory different compared to a "high memory" settings of 20GB. Specifically, do the same settings apply to lower memory settings as they would to higher?

[DanVanAtta]

To refine/clarify the questions a bit:

• can you count the number of partial and full GCs and their average + max durations?
• do you see a shift in GC performance when throttled down to 8GB (the 20GB might be a sweet spot because you remove all full GCs while still keep the heap small enough that partial GCs are not slow themselves)
• What performance differences do you see at 8GB compared to 20GB, and can you pinpoint why 20GB is performing better?
• Are the "best" settings the same between 20GB and 8GB heap sizes, or do the other tunings need to change?

[M0ON-MAN]

Thanks for the response and appreciate the input. I intend to test all this as i find time between work, certification exam prep, family and life... lol! i will post as/when able. I do want to complete(some components are backorder or held up due to tariffs) the 128GB ram, Framework Desktop machine so i can run the game in a VM and get a deeper dive/understanding of how java 26 and java 21 behave and perform differently using a save from a 600+ round/turn.