So time to reduce latency using JVM option, or history about random vs urandom differences

 Hi,

 

Sometimes, I need to reduce latency of Java app, especially, of Atlassian Suite. 

As we know Jira, use UUID generation for many approaches, like integration with other Atlassian products. as example, I made grepping for the 7.13.x release sources for the UUID references

image.png

So we can optimize a small option in the java.security, 

which is located in Java 8.

$JAVA_HOME/lib/security/java.security

for Java 11: 

$JAVA_HOME/conf/security/java.security

 

We can change  to 

securerandom.source=file:/dev/urandom

instead of 

securerandom.source=file:/dev/random

I do recommend you to read pages linked in References.

 

Benchmarks

Before making change, let’s do a few  benchmarks. 

I used that one (https://github.com/ajbrown/uuid-generator-benchmark)

on the 

Linux version 3.10.0-1127.el7.x86_64 (mockbuild@x86-034.build.eng.bos.redhat.com) (gcc version 4.8.5 20150623 (Red Hat 4.8.5-39) (GCC) ) #1 SMP Tue Feb 18 16:39:12 EST 2020

 

Results:

/dev/random - jdk8u242-b08
# Run complete. Total time: 00:02:42

Benchmark                           Mode Cnt Score Error Units
EAIOBenchmark.timeAndEthernetBased    ss 100 83.127 ± 4.634 ms/op
JUGBenchmark.nameBased                ss 100 93.167 ± 8.612 ms/op
JUGBenchmark.random                   ss 100 198.932 ± 12.170 ms/op
JUGBenchmark.timeAndEthernetBased     ss 100 93.234 ± 7.041 ms/op
JUGBenchmark.timeBased                ss 100 92.625 ± 8.496 ms/op
JavaUUIDBenchmark.nameBased           ss 100 206.336 ± 36.419 ms/op
JavaUUIDBenchmark.random              ss 100 288.553 ± 25.660 ms/op

 

/dev/random - openjdk 11.0.6
# Run complete. Total time: 00:02:29

Benchmark                          Mode Cnt Score Error Units
JUGBenchmark.nameBased               ss 100 96.258 ± 13.504 ms/op
JUGBenchmark.random                  ss 100 199.150 ± 16.954 ms/op
JUGBenchmark.timeAndEthernetBased    ss 100 88.397 ± 7.082 ms/op
JUGBenchmark.timeBased               ss 100 87.346 ± 7.794 ms/op
JavaUUIDBenchmark.nameBased          ss 100 137.397 ± 29.385 ms/op
JavaUUIDBenchmark.random             ss 100 227.381 ± 20.810 ms/op

 

/dev/urandom - jdk8u242-b08
# Run complete. Total time: 00:02:38

Benchmark                           Mode Cnt Score Error Units
EAIOBenchmark.timeAndEthernetBased    ss 100 86.825 ± 5.337 ms/op
JUGBenchmark.nameBased                ss 100 92.620 ± 8.273 ms/op
JUGBenchmark.random                   ss 100 196.341 ± 13.258 ms/op
JUGBenchmark.timeAndEthernetBased     ss 100 90.217 ± 8.017 ms/op
JUGBenchmark.timeBased                ss 100 85.742 ± 6.362 ms/op
JavaUUIDBenchmark.nameBased           ss 100 194.917 ± 31.265 ms/op
JavaUUIDBenchmark.random              ss 100 276.298 ± 26.933 ms/op

 

 /dev/urandom - openjdk 11.0.6
# Run complete. Total time: 00:02:19

Benchmark                          Mode Cnt Score Error Units
JUGBenchmark.nameBased               ss 100 88.819 ± 10.755 ms/op
JUGBenchmark.random                  ss 100 179.651 ± 12.303 ms/op
JUGBenchmark.timeAndEthernetBased    ss 100 81.948 ± 4.981 ms/op
JUGBenchmark.timeBased               ss 100 82.964 ± 4.402 ms/op
JavaUUIDBenchmark.nameBased          ss 100 124.635 ± 25.655 ms/op
JavaUUIDBenchmark.random             ss 100 210.720 ± 16.614 ms/op

image.png

 

Conclusion

The result of benchmarks shows reduced latency results, but not a big, because I did microbenchmark, which is shows exact function for the big system you can see a good improvements.

 

I like philosophy of storing config in the environment or on related CI/CD tools.

Hence you can use the JVM parameter in setenv.sh (setenv.bat) like this

-Djava.security.egd=file:///dev/urandom

 

References:

https://www.ibm.com/support/pages/slow-performance-linux-vm-701-and-prior-releases

https://quotidian-ennui.github.io/blog/2012/06/15/slow-java-crypto-performance-on-linux/

 

Cheers,

Gonchik Tsymzhitov

Comments

Popular posts from this blog

How only 2 parameters of PostgreSQL reduced anomaly of Jira Data Center nodes

Atlassian Community, let's collaborate and provide stats to vendors about our SQL index usage

How do you analyze GC logs, thread dumps and head dumps?