Eclipse TPTP

TPTP Heap and Thread ProfilersHigh-level DesignRev 1.0

TPTP Heap and Thread ProfilersHigh-level DesignRev 1.0

Asaf Yaffe

July, 2006

Eclipse TPTP

Heap ProfilerHeap Profiler

Eclipse TPTP

Profiler RequirementsProfiler Requirements• Track object allocations and de-allocations

• Track GC events

• Produce heap dumps on demand

• Generate XML trace compatible with the JVMPI Heap Profiler

Eclipse TPTP

Martini RequirementsMartini Requirements• Generate the following events

– Object Alloc– Object Free– GC Start– GC Finish

• Support the following data requests– Get Object information (from an object id)– Get Class information (from a class id)– Get Reference graph for living objects (heap dump)

Eclipse TPTP

Martini Heap Events OverviewMartini Heap Events Overview• Martini uses intra-function instrumentation for generating Object Alloc

events• For all reference types except arrays, the Object Alloc event is generated

by instrumenting the constructor of the java.lang.Object class– The instrumentation code passes the newly allocated object reference to a

Java static recorder function HeapProxy.ObjectAlloc which calls back into JPI

– JPI tags the allocated object (via JVMTI) in order for the JVM to identify the object in the Object Free event

• communicated by the JVM via JVMTI

– This technique is sub-optimal but easy to implement• Selectivity is implemented on Object Alloc event rather than during instrumentation• May be optimized in future slices

• For array types, the Object Alloc event is generated by instrumenting array allocation sites (i.e., after newarray, anewarray, multianewarray)

• The instrumentation activity takes place during CLASS_LOAD_HOOK event of the java.lang.Object class or when heap events are dynamically enabled by the tool

Eclipse TPTP

The HeapProxy ClassThe HeapProxy Class• Martini has a helper HeapProxy Java class

– Resides in a directory that is added to the JVM boot CLASSPATH during JPI initialization

• The Problem: the HeapProxy recorder methods are defined as “native” and are bound by JPI only during the JVM_INIT event (because JNI services are not available before JVM_INIT)– This is problematic since for some system classes, CLASS_LOAD_HOOK

events are sent before JVM_INIT

– Inserting calls to native methods before they are bound cause the JVM to crash, even if these methods are never executed (just referencing these methods in the code is enough to crash the JVM)

• The Solution: HeapProxy has wrapper functions which are always called by instrumentation code that is inserted before JVM_INIT– Those functions check whether the JVM has already been initialized and call

the native ObjectAlloc functions only after the JVM has been initialized• i.e. object allocations before JVM_INIT after not communicated to the heap profiler

Eclipse TPTP

Object Alloc EventObject Alloc Event• Event Source

– Regular Objects: BCI of java.lang.Object.<init>– Arrays: BCI after callsites of newarray, anewarray, multianewarray– Internal VM Allocations: JVMTI_EVENT_VM_OBJECT_ALLOC

• Data Items– Object ID– Class ID– Thread id (of the allocating thread)– Object Size– Whether this object is an array

• Selectivity– By package/class

Eclipse TPTP

Object Free EventObject Free Event• Event Source

– JVMTI_EVENT_OBJECT_FREE for tagged objects

• Data Items– Object ID (easy - JVMTI)– All other information is optional

• Can also be looked-up using a data request

• Selectivity– By fully qualified class name– Defined at Object Allocation time for both Object Alloc and Free

events

Eclipse TPTP

Garbage Collection StartedGarbage Collection Started• Event Source

– JVMTI_EVENT_GARBAGE_COLLECTION_START

• Data Items– None

• Selectivity– None

• Notes– Generated only for “stop-the-world” collections. Other collections are

not reported.

Eclipse TPTP

Garbage Collection FinishedGarbage Collection Finished• Event Source

– JVMTI_EVENT_GARBAGE_COLLECTION_FINISH

• Data Items– None

• Selectivity– None

• Notes– See Garbage Collection Started

Eclipse TPTP

Heap Dump/Object ReferencesHeap Dump/Object References• TBD…

Eclipse TPTP

Open IssuesOpen Issues• Heap Dump data

– Need more info about the following XML fragments and their attributes:

• heapDumpDef• gcRoot• objectReference

• gcFinish element– Are the following attributes used by the viewers:

usedObjects, usedObjSpace, totalObjSpace

Eclipse TPTP

Thread ProfilerThread Profiler

Eclipse TPTP

Profiler RequirementsProfiler Requirements• Track Thread start and end events

• Track monitor events (critical sections, wait)

• Generate XML trace compatible with the JVMPI implementation

Eclipse TPTP

Martini RequirementsMartini Requirements• Generate the following events

– Thread Start– Thread End– Contended Monitor Enter– Contended Monitor Entered– Monitor Wait– Monitor Waited

• Support the following data requests– Get Object Monitor Usage information (from an object id)– Get Thread information (from a thread id)

Eclipse TPTP

Thread Start EventThread Start Event• Event Source

– JVMTI

• Data Items– Unique Thread ID– Thread Name– Parent Name– Group Name

• Selectivity– None

Eclipse TPTP

Thread End EventThread End Event• Event Source

– JVMTI

• Data Items– Thread ID

• Selectivity– None

Eclipse TPTP

Contended Monitor Enter EventContended Monitor Enter Event• Event Source

– JVMTI

• Data Items– Unique Monitor ID– The waiting thread id– The id of the thread owning the monitor

• Notes– Other monitor information can be obtained with a data request:

• For contended monitors: a list of threads waiting for the monitor• For “wait” monitors: a list of threads waiting to be notified

Eclipse TPTP

Contended Monitor Entered EventContended Monitor Entered Event• Event Source

– JVMTI

• Data Items– Monitor ID– The id of the thread that entered the monitor

• Selectivity– None

Eclipse TPTP

Monitor Wait EventMonitor Wait Event• Event Source

– JVMTI

• Data Items– Monitor ID– The id of the waiting thread– Wait timeout - the value passed to the obj.Wait() method

• Selectivity– None

Eclipse TPTP

Monitor Waited EventMonitor Waited Event• Event Source

– JVMTI

• Data Items– Monitor ID– The id of the waiting thread

• Selectivity– None

• Notes– The <monWaited> output element includes the “timeout” attribute that

specifies the number of milliseconds the thread had to wait for the monitor.

– This data item is not available in JVMTI so the profiler will have to calculate it manually

Eclipse TPTP

BackupBackup

Eclipse TPTP

HeapProxy Java ClassHeapProxy Java Classpublic class HeapProxy {

private static int m_bJVMInit = 0;

public static void JVMInit() { m_bJVMInit = 1; }

public static native void ObjectAlloc(java.lang.Object obj);

public static void EarlyObjectAlloc(java.lang.Object obj) {

if (m_bJVMInit != 0) ObjectAlloc(obj); }

public static native void ArrayAlloc(java.lang.Object arr);

public static void EarlyArrayAlloc(java.lang.Object arr) {

if (m_bJVMInit != 0) ArrayAlloc(obj);}

}

Eclipse TPTP

Instrumenting java.lang.Object ConstructorInstrumenting java.lang.Object Constructor

Before After

public Object()

{

}

public Object(){ HeapProxy.EarlyObjectAlloc(this);}

Best seen in Slide Show mode

Martini inserts a call to HeapProxy.EarlyObjectAlloc, because the insertion will always happen before the VM is initialized, and we cannot use native methods because they were not bound yet.

Eclipse TPTP

Instrumenting Array Allocation Callsites[case 1 – before JVM_INIT event received]

Instrumenting Array Allocation Callsites[case 1 – before JVM_INIT event received]

Before After

void someMethod()

{

…

…

int[] intArray =

new int[20];

…

…

}

Best seen in Slide Show mode

void someMethod()

{

…

…

int[] intArray =

new int[20];

HeapProxy.EarlyArrayAlloc(intArray);

…

…

}

Martini inserts a call to HeapProxy.EarlyArrayAlloc. Since the JVM has not been initialized yet, we cannot use native methods because they were not bound yet.

Eclipse TPTP

Instrumenting Array Allocation Callsites[case 2 – after JVM_INIT event received]Instrumenting Array Allocation Callsites

[case 2 – after JVM_INIT event received]

Before After

void someMethod()

{

…

…

int[] intArray =

new int[20];

…

…

}

Best seen in Slide Show mode

void someMethod()

{

…

…

int[] intArray =

new int[20];

HeapProxy.ArrayAlloc(intArray);

…

…

}

Martini inserts a call to HeapProxy.ArrayAlloc. Since the JVM has already initialized, we can safely use native methods because they were already bounded.

Eclipse TPTP

Heap Instrumentation OverviewHeap Instrumentation Overview

• Instrumenting the java.lang.Object constructor:– Add references to the HeapProxy class and the

HeapProxy.EarlyObjectAlloc function to the constant pool– Modify the method’s code attribute and inject code to call the

HeapProxy.EarlyObjectAlloc function.

• Instrumenting array allocation sites in a Java class:– Add references to the HeapProxy class and the HeapProxy.ArrayAlloc

or HeapProxy.EarlyObjectAlloc functions to the constant pool– Iterate over the class methods, and for each method that can be

instrumented (i.e. has a “code” attribute)• Search for any of the following array allocation instructions: newarray,

anewarray, multianewarray.

• If found, modify the method’s code attribute and inject code after each allocation site to call the HeapProxy.ArrayAlloc function.

Eclipse TPTP

java.lang.Object Constructor Instrumentation Detailsjava.lang.Object Constructor Instrumentation Details

return

.end

.start

Original method code

Legendpseudo-instructions

original instructions

new instructions

.start

inst1

sipush

Instrumented method code

.start

aload_0

invokestatic EarlyObjectAlloc()

.end

return

The ‘this’ pointer of the newly allocated object is stored in local variable 0. Load it to the operand stack so it can be passed to the EarlyObjectAlloc static function

Eclipse TPTP

Array Allocation Site Instrumentation DetailsArray Allocation Site Instrumentation Details

.end

.start

Original method code

Legendpseudo-instructions

original instructions

new instructions

.start

inst1

sipush

Instrumented method code.start

dup

invokestatic [Early]ArrayAlloc()

.end

inst

newarray

inst

inst

inst

inst

inst

newarray

inst

inst

After the ‘newarray’ instruction, a reference to the newly allocated array is on the operand stack. The reference is duplicated and sent to the ArrayAlloc or EarlyArrayAlloc static function