I have a microservice (springboot) inside a docker container. I am also passing a long list of JVM flags so that the program runs on that VM environment to test the latency values with each flag combination.
I am using this command to start the container:
docker run --rm --name factorialorialContainer -p 8080:8080 -e JAVA_OPTIONS="$(cat /Users/sulekahelmini/Documents/fyp/fyp_work/MLscripts/flags.txt)" suleka96/factorial:latest
The flags.txt looks like this:
-XX:-ResizePLAB -XX:+ResizeOldPLAB -XX:+AlwaysPreTouch -XX:-ParallelRefProcEnabled -XX:-ParallelRefProcBalancingEnabled -XX:-UseTLAB -XX:-ResizeTLAB -XX:+ZeroTLAB -XX:+FastTLABRefill -XX:-UseAutoGCSelectPolicy -XX:-UseAdaptiveSizePolicy -XX:+UsePSAdaptiveSurvivorSizePolicy -XX:-UseAdaptiveGenerationSizePolicyAtMinorCollection -XX:+UseAdaptiveGenerationSizePolicyAtMajorCollection -XX:-UseAdaptiveSizePolicyWithSystemGC -XX:+UseAdaptiveGCBoundary -XX:+UseAdaptiveSizePolicyFootprintGoal -XX:YoungPLABSize=4354 -XX:OldPLABSize=1141 -XX:GCTaskTimeStampEntries=255 -XX:TargetPLABWastePct=5 -XX:PLABWeight=73 -XX:OldPLABWeight=21 -XX:MarkStackSize=5927008 -XX:MarkStackSizeMax=579749070 -XX:RefDiscoveryPolicy=1 -XX:InitiatingHeapOccupancyPercent=49 -XX:ErgoHeapSizeLimit=0 -XX:MaxRAMFraction=3 -XX:MinRAMFraction=3 -XX:InitialRAMFraction=74 -XX:AutoGCSelectPauseMillis=2934 -XX:AdaptiveSizeThroughPutPolicy=0 -XX:AdaptiveSizePausePolicy=0 -XX:AdaptiveSizePolicyInitializingSteps=24 -XX:AdaptiveSizePolicyOutputInterval=4060 -XX:AdaptiveSizePolicyWeight=11 -XX:AdaptiveTimeWeight=69 -XX:PausePadding=39200 -XX:PromotedPadding=3 -XX:SurvivorPadding=1 -XX:ThresholdTolerance=11 -XX:AdaptiveSizePolicyCollectionCostMargin=68 -XX:YoungGenerationSizeIncrement=28 -XX:YoungGenerationSizeSupplement=81 -XX:YoungGenerationSizeSupplementDecay=9 -XX:TenuredGenerationSizeIncrement=48 -XX:TenuredGenerationSizeSupplement=63 -XX:TenuredGenerationSizeSupplementDecay=2 -XX:MaxGCPauseMillis=16157174462788231168 -XX:GCPauseIntervalMillis=3009 -XX:GCTimeRatio=77 -XX:AdaptiveSizeDecrementScaleFactor=4 -XX:AdaptiveSizeMajorGCDecayTimeScale=9 -XX:MinSurvivorRatio=4 -XX:InitialSurvivorRatio=4 -XX:BaseFootPrintEstimate=463170010 -XX:GCHeapFreeLimit=4 -XX:ProcessDistributionStride=4 -Xms338224019 -Xmx1103493303
when I run the above-mentioned docker command I get the below error:
#
A fatal error has been detected by the Java Runtime Environment:
#
SIGSEGV (0xb) at pc=0x00007f9b014fe3c3, pid=1, tid=0x00007f9b10633b10
#
JRE version: OpenJDK Runtime Environment (8.0_212-b04) (build 1.8.0_212-b04)
Java VM: OpenJDK 64-Bit Server VM (25.212-b04 mixed mode linux-amd64 compressed oops)
Derivative: IcedTea 3.12.0
Distribution: Custom build (Sat May 4 17:33:35 UTC 2019)
Problematic frame:
j java.net.URLStreamHandler.toExternalForm(Ljava/net/URL;)Ljava/lang/String;+88
#
Failed to write core dump. Core dumps have been disabled. To enable core dumping, try "ulimit -c unlimited" before starting Java again
#
An error report file with more information is saved as:
/usr/src/factorial/hs_err_pid1.log Compiled method (c1) 1054 256 1 java.net.URL::getRef (5 bytes) total in heap
[0x00007f9b0166bad0,0x00007f9b0166bd60] = 656 relocation
[0x00007f9b0166bbf8,0x00007f9b0166bc20] = 40 main code
[0x00007f9b0166bc20,0x00007f9b0166bca0] = 128 stub code
[0x00007f9b0166bca0,0x00007f9b0166bd30] = 144 scopes data
[0x00007f9b0166bd30,0x00007f9b0166bd38] = 8 scopes pcs
[0x00007f9b0166bd38,0x00007f9b0166bd58] = 32 dependencies
[0x00007f9b0166bd58,0x00007f9b0166bd60] = 8 #If you would like to submit a bug report, please include
instructions on how to reproduce the bug and visit:
https://icedtea.classpath.org/bugzilla
#
What am I doing wrong here?
It's -XX:+ZeroTLAB option which is broken. Just don't use this flag.
I guess a random option generator is involved here for the purpose of performance tuning. While the idea itself is fine, and there are known cases1,2 when Machine Learning with Bayesian Optimization helped to find better values of JVM options, the key point here is to include in the experiments only those options which meaning and implications you understand well.
Also the experiments should take care of the reasonable ranges of each option, as well as the mutual relationship between connected options.
However, the above option list looks totally random and does not make much sense (maybe, except for JVM testing). No surprise that such configuration may produce unpredictable results, including a JVM crash.
1 Automated Tuning of the JVM with Bayesian Optimization
2 Performance tuning Twitter services with Graal and Machine Learning