JVM Performance Tuning Tips

->     Problems
–     Same as client side, default collector used most of the time

–     heaps are sized    bigger

–     stop the world collection

->     See
–     larger pauses as heap sizes are bigger

–     GC frequency dependent on load

–     CPUs are idle during collection

–     scalability problems


Performance Tuning

->     Manual tuning by application modeling
–     What is application modeling

–     GC Portal

–     General Tuning Tips

->     Automatic tuning
–     Ergonomics

Application    Modeling From A GC Perspective

Application    Modeling    and Performance Analysis

->     What is it->

->     Recommendations based on the model

->     Empirical modeling

->     Theoretical projections

Application Modeling And Performance Tuning: GC Perspective

->    Goal : remove the unpredictable behavior of an application

->    Construct a mathematical model by mining the verbose GC log files

->    The model takes into account

–     Incoming load information

–     Data in verbosegc log files

The Model
Incoming load information

->    Transaction Rate (Allocation Rate)

->     Active Transaction Duration
–     Lifetimes of short and long lived data

->     Size of objects per transaction

The Model (1)
Data in the verbose GC log files
->    GC    pauses

–     Young and Old generation pauses

–     Time to start-stop application threads

–     Application time

->     GC frequency
–     Young and old generation periodicity

->     Rate of allocation/promotion of objects

->     Direct allocation of objects in old generation

The Model (2)
Data in the verbosegc log files (Contd.)

->     Total
–     GC time, Application time

–     Objects promoted

–     Garbage collected

->     Heap sizes
–     Size of Young generation (Eden, Semi-Space)

–     Size of    Old generation

–     Initial and Final Size of old generation

–     Size of Permanent generation

–     Average occupancy, and heap thresholds for GC

Verbosegc Log    Sample
0.740905: [GC {Heap before GC invocations=8:
def new generation    total 1536K, used 1055K [0xf2c00000, 0xf2e00000, 0xf2e00000)
eden space        1024K, 99% used [0xf2c00000, 0xf2cfdfe0, 0xf2d00000) from space    512K,    7% used [0xf2d00000, 0xf2d09c50, 0xf2d80000) to    space    512K,    0% used [0xf2d80000, 0xf2d80000, 0xf2e00000)
concurrent mark-sweep generation total 59392K, used 540K [0xf2e00000, 0xf6800000, 0xf6800000)
concurrent-mark-sweep perm gen total 4096K, used 1158K [0xf6800000, 0xf6c00000, 0xfa800000)

0.741773: [DefNew Desired survivor size 262144 bytes, new threshold 1 (max 31)
age    1:    280048 bytes,        280048 total age    2:     40016 bytes,    320064 total
: 1055K->312K(1536K), 0.0048282 secs] 1595K->853K(60928K)

Heap after GC invocations=9:Heap
def new generation    total 1536K, used 312K [0xf2c00000, 0xf2e00000, 0xf2e00000)
eden space        1024K,    0% used [0xf2c00000, 0xf2c00000, 0xf2d00000 ) from space 512K, 61% used [0xf2d80000, 0xf2dce240, 0xf2e00000) to    space    512K,    0% used [0xf2d00000, 0xf2d00000, 0xf2d80000 )
concurrent mark-sweep generation total 59392K, used 540K [0xf2e00000, 0xf6800000, 0xf6800000)
concurrent-mark-sweep perm gen total 4096K, used 1158K [0xf6800000, 0xf6c00000, 0xfa800000)} , 0.0063803 secs]

Data    Calculated

->     GC sequential overhead (Directly related to application throughput)
->     GC concurrent overhead
->     Average size of objects
->     Active data duration (long and short term objects)
->     Actual throughput
->     Application efficiency
->     Speedup (Amdahl’s law)
->     % CPU utilization
->     Memory Leak detection

General Recommendations based    on    the model (1)

->    General JVM Tuning and Sizing methodology
–     Size of old generation = Call rate * active call duration * long lived data/call
–     Size of young generation = Call rate * expected periodicity of GC * short lived data/call
->     for desired pause and frequency
->     Reduce GC pauses
->    Reduce GC sequential overhead
General Recommendations
Based    On    The Model (2)

->    Size the young and old generation heaps to handle a given load
->     Detect memory leaks
->     Choice of collector
->     Choice of the different JVM options and switches

Empirical Modeling

->     Rank the Application runs based on data analyzed from the verbosegc logs
->     Choose the optimum JVM environment based on criteria:

–     Heap sizes
–     No. of Processors
–     GC sequential overhead
–     GC concurrent overhead, etc.
–     Application efficiency

Theoretical Projections    For Tuning    Based    On    The Model

->    “What-if” scenarios could be tried

–     How GC behavior changes with change in Application and JVM parameters

->    “What-if” input parameters include:

–     Size of young generation
–     Size of old generation
–     Request rate/Load
–     Garbage/request
–     No. of processors

Theoretical Projections For Tuning

->    Projection output shows :
–     what could be the

->     GC pause (latency)
->     GC frequency
->     GC sequential load (bandwidth)
->     % CPU utilization, Speedup
->     Application efficiency
->     Allocation rate, Promotion rate
->     Size and duration of Short lived data
->     Size and duration of Long lived data

GC Portal

Enables as a service, Application Modeling and Performance Tuning from GC perspective.
->    Implemented in J2EE
->    Allows developers to submit log files, and analyze application behavior
->    Portal can be used to performance tune, and size application to run optimally under lean, peak, and burst conditions

SnapShot from the GC Portal
GC Portal

GC Portal

->     Plots and displays graphically GC behavior over time. Parameters include:

–     GC pauses (Max. and Average)

–     GC frequency

–     GC sequential load

–     GC concurrent load

–     Garbage Allocation rate

–     Garbage Promotion rate

Snapshot from the GC portal Graphical Engine
GC portal Graphical Engine

GC Portal
->     Provides General Recommendations

->    Projections for sizing and tuning via
“what-if” scenarios
->     Empirical modeling
Snapshot from the GC portal
What-if scenarios
GC portal

GC General Tuning    Tips

Reducing    Collection    Times
->    Use -Xconcgc for low pause applications
->    Use -XX:+AggressiveHeap for throughput applications
–     Use -XX:+PrintCommandLine to see AggressiveOptions, and use this to tune further

->    Size Permanent Generation
->    Reduce pooled objects
->    Using NIO
->    Avoid System.gc() and distributed RMI GC
–     Use -XX:+DisableExplicitGC

->    Making immutables, mutables
–     String -> String Buffer for String manipulation, and maybe storage
->     Avoiding old generation undersized heaps
–     Reduces collection time, but leads to lot of other problems like fragmentation, triggers Full GC

Reducing    Frequency Of GC

->    Frequency of a collection is dependent on
–     Size of young and old generations
–     Incoming load
–     Object life time
->     Increase young generation to decrease frequency of collection but this will increase pause
–     Choose a size where pause is tolerable

->     Increase in load will fill up the heap faster so increases collection frequency
–     Increase heap to reduce frequency

->    Increase in lifetime of objects increases frequency as live objects take space
–     Keep live objects to the needed minimum

Sizing    The Heap

->     Heap size influences the following
–     GC frequency and collection times
–     Number of short and long term objects
–     Fragmentation and locality problems

->     Undersized heap with concurrent collector
–     leads to Full GCs with increase in load

–     Fragmentation problems

->     Oversized heap
–     leads to increased collection times

–     locality problems (smear problem)

–     Use ISM and variable page sizes to reduce smear problem

->     Size heap to handle peak and burst loads

Improving Execution Efficiency

->    GC Portal computes execution efficiency

->    Efficiency calculated using Amdahl’s law

->    Translates to CPU utilization

->    Higher this value the better

->    Increase efficiency by reducing serial parts

–     Reducing GC pause & frequency

–     Reducing long term objects and increasing short term objects

–     Creating only needed objects like using NIO, mutables

–     Avoiding Full GC, For e.g. RMI DGC, undersized heaps

–     Choosing optimum heap size to reduce smear effect

Other    Ways To Improve    Performance On Solaris

->     Using the Solaris RT (real-time) scheduling class

->     Using the alternate thread library (/usr/lib/lwp)
– default thread library on Solaris 9

->     Using hires_tick to change clock resolution

->     Using processor sets

->     Binding process to a CPU

->     Turning off    interrupts

->     Modifying the dispatch table

->     Use large page sizes

->     Use multi-threaded malloc library

Automatic Tuning In    J2SE 1.5


->     What is ergonomics->

->     Why do it->

->     When is it used->

->     What does it do->

->     How does it work->

What Is    Ergonomics->

->     JVM™ automatically selects
–     Compiler
–     Garbage collector
–     Heap size
->     User specifies behavior

->     GC dynamically does tuning
–     AKA GC ergonomics
Why Do    Ergonomics->

->     Better Performance
–     Hand tuned performance is good

->     Ease of Use
–     Hand tuning is hard

->     Better Resource Usage
–     Use what you need

When    Is Ergonomics    Used->

->     Server class machines
–     2 CPUs, 2 Gbytes

->     Exceptions
–     Microsoft Windows ia32

What Does Ergonomics    Do->

->     Server compiler

->     Parallel GC collector

->     Maximum heap
–     Smaller of

->  ¼ physical memory

->  1 Gbyte

->     Initial heap
–     Smaller of

->  1/64 physical memory

->  1 Gbyte

What is    GC Ergonomics->

->     User specifies
–     Maximum pause time goal

–     Throughput goal

–     Assumes minimum footprint goal

->     GC tunes
–     Young generation size

–     Old generation size

–     Survivor space sizes

–     Tenuring threshold

Why Do    GC Ergonomics->

->     Common complaints
–     Pauses are too long

–     GC is too frequent

->     Solution
–     Hand tune the GC
What Does GC Ergonomics
->     Goals, not guarantees

->     User specified behavior
–     Maximum pause time goal

->  Reduce size of generation

–     Throughput goal

->  Increase size of generations

–     Minimum footprint

->  Reduce size of generations

->     Again, goals, not guarantees

Ergonomics    Usage
->     Use -XX:+UseParallelGC with the below options

->     Throughput Goal
–     -XX:GCTimeRatio=nnn

->    The ratio of GC time to application time

->        1 / (1 + nnn) where nnn is a value to obtain the percentage GC time vs application time. E.g. Nnn =
19,    GC time 5% of application time

->     Pause Time Goal
–     -XX:MaxGCPauseMillis=nnn

->  An hint to the JVM to keep the pauses below this value

Ergonomics    Strategy

->     Use throughput strategy, and set desired throughput

->     Change maximum heap size if throughput cannot be achieved

->     If throughput goal is achieved, set pause time goal, if pauses are high

JVM    Monitoring & Management in    J2SE 1.5

Java Monitoring    and    Management API
->     Provides a way to manage and monitor a JVM
–     Information about loaded classes and threads

–     Memory usage

–     Garbage collection statistics

–     Low memory detection & thresholds

->     Provides monitoring utilities
–     jconsole

–     jstat

Java Monitoring    and    Management API
->     Provides MBeans
–     GarbageCollectorMXBean

–     MemoryManagerMXBean

–     MemoryMXBean

–     MemoryPoolMXBean

–     Other MBeans

->     MBeans can be accessed through
–     jconsole

->  jconsole jvmpid

jconsole – GarbageCollection

jconsole - GarbageCollection
jconsole – Memory Usage

jconsole - Memory Usage


->     An utility to obtain JVM statistics dynamically
–     Compiler statistics

–     Class loader statistics

–     GC statistics

->     GC statistics include
–     cause of GC
–     generation information

->  capacity
->  utilization
jstat Usage

->     jstat -gc jvmid
–     Provides statistics on the behavior of the garbage collected heap

->     jstat -gcutil jvmid
–     Provides a concise summary of garbage collection statistics.

->     jstat -gcutil 21891

–    S0        S1       E       O         P     YGC    YGCT    FGC    FGCT    GCT

–    12.44    0.00    27.20    9.49    96.70    78    0.176    5    0.495    0.672

–    12.44    0.00    62.16    9.49    96.70    78    0.176    5    0.495    0.672

–    12.44    0.00    83.97    9.49    96.70    78    0.176    5    0.495    0.672

–    0.00    7.74    0.00    9.51    96.70    79    0.177    5    0.495    0.673


->    Introduced low pause and throughput collectors

->    Performance problems seen with garbage collection

->    Improving performance using manual and automatic tuning

->    Introduction to the new monitoring & management API


->    http://java.sun.com/docs/hotspot/index.html

->     http://java.sun.com/docs/hotspot/gc1.4.2/

->     http://developers.sun.com/techtopics/mobility/midp/articles/garbagecollection2/

->     http://http://java.sun.com/developer/technicalArticles/Programming/turbo/

->     http://java.sun.com/docs/hotspot/VMOptions.html

->     http://sdc.sun.com/gcportal/

->     http://java.sun.com/developer/technicalArticles/Programming/GCPortal/index.html

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.