How to easy decrease of heap size for your server product?

[#1]

Hi, community! 

I would like to share my experience using the String deduplication functionality in JVM. 

A suggestion was based on community questions and GC logs provided by members, where I have not seen that feature on JVM parameters. 

Disclaimer,  as usual, everything should be tested in the test-environment first.

My use case was started from the next situation, if you have a system where you can not extend RAM easier and don't want to use as a swap file.  

Well let's start, as you know that a string deduplication feature added into Java 8 update 20 (it is super old release, 2014).

All technical info you can in JEP-192 - http://openjdk.java.net/jeps/192

You can enabled in G1 GC with next parameters:

-XX:+UseG1GC -XX:+UseStringDeduplication -XX:+PrintStringDeduplicationStatistics

• UseStringDeduplication (bool) -  Enable string deduplication

• PrintStringDeduplicationStatistics (bool) - Print detailed deduplication statistics

• StringDeduplicationAgeThreshold (uintx) - String objects reaching this age will be considered candidates for deduplication

Let's see the results on monitoring:

The yellow line is non-heap memory usage, the green is heap memory. 

In this case, I especially removed the value of GB. Because that optimisation is really individual for every instance. Because we are using 3rd-party plugins and own apps. Where all those parameters reflected to value and also, some plugins (apps) starting to be better after implementing the DC compatibility requirements. (Thanks for that Atlassian). 

But the  last statistics you can read in this stats output, reading that info I decreased in our 3rd party app a String usage.

2019-03-03T22:32:49.221+0100: 843349.318: [GC concurrent-string-deduplication, 15.6M->82.4K(15.5M), avg 78.5%, 0.1523212 secs]
[Last Exec: 0.1523212 secs, Idle: 11.6095757 secs, Blocked: 2/0.0201580 secs]
[Inspected: 184339]
[Skipped: 0( 0.0%)]
[Hashed: 136198( 73.9%)]
[Known: 30( 0.0%)]
[New: 184309(100.0%) 15.6M]
[Deduplicated: 182894( 99.2%) 15.5M( 99.5%)]
[Young: 182894(100.0%) 15.5M(100.0%)]
[Old: 0( 0.0%) 0.0B( 0.0%)]
[Total Exec: 22129/365.6814278 secs, Idle: 22129/842910.6783849 secs, Blocked: 14649/68.0063332 secs]
[Inspected: 434930188]
[Skipped: 2( 0.0%)]
[Hashed: 270134906( 62.1%)]
[Known: 2133448( 0.5%)]
[New: 432796738( 99.5%) 27.6G]
[Deduplicated: 405747212( 93.8%) 21.6G( 78.5%)]
[Young: 213685766( 52.7%) 12.4G( 57.5%)]
[Old: 192061446( 47.3%) 9423.7M( 42.5%)]
[Table]
[Memory Usage: 258.9M]
[Size: 8388608, Min: 1024, Max: 16777216]
[Entries: 8146001, Load: 97.1%, Cached: 371239, Added: 28627150, Removed: 20481149]
[Resize Count: 13, Shrink Threshold: 5592405(66.7%), Grow Threshold: 16777216(200.0%)]
[Rehash Count: 0, Rehash Threshold: 120, Hash Seed: 0x0]
[Age Threshold: 3]
[Queue]
[Dropped: 0]
[GC concurrent-string-deduplication, deleted 1 entries, 0.0000053 secs]

Well, I hope you have concerns related to time for execute that functionality:

Where we can see that function works often. but it is interesting how much need to complete 1 time review string deduplication.

 

As you see it fastest phase of G1 GC.

Let's see the situation of with CPU, because we need to pay to CPU :).

In my situation, everything is ok. Also, I see my CPU is wasting time ;)

Conclusion:

If you want to balance CPU and RAM usage for your JVM, by string deduplication feel free to use it. Because it is an easy to win.  

I hope that info was interesting for you. If It is, I will share the next some parameters I use for one of my env.

Cheers,

Gonchik Tsymzhitov

[Marty]

Hi Gonchik,

this is interesting! Anything we can do to squeeze better performance out of the JVM, the better.

My only question is - have you observed any downside from switching to the G1 Garbage Collector?

I note that for jdk9 onwards the G1GC is the default Garbage Collector, but I see that Java 8 is the latest supported Java for Jira / Confluence Server.

[Gonchik Tsymzhitov]

Hi @Marty , 

Sorry for late answer, (I have not seen that  comment :( in notifications )

I have not met with downside to the switching from Parallel GC. I met once around 2 years ago, with switching from CMS, but that instance was super deeply configured for CMS parameters. BTW, after playing around G1 parameters (https://www.oracle.com/technetwork/articles/java/g1gc-1984535.html ) mostly I used 3-4 parameters hence everything was pretty equaled as CMS on that Jira instance.

BTW, keep in your mind from Java 9 CMS was deprecated

https://community.atlassian.com/t5/Data-Center-articles/Which-GC-strategy-do-you-prefer-for-your-app/ba-p/1052212

[Gonchik Tsymzhitov]

@Marty Feel free to comment on that story related G1GC and FullGC situation:

https://community.atlassian.com/t5/Data-Center-articles/Why-do-you-need-to-use-the-JDK-11-instead-of-8/ba-p/1666278