runtime innovation - nextgen ninja hacking of the jvm, by ryan sciampacone

© 2013 IBM Corporation

Runtime Innovation: NextGen Ninja Hacking of the JVM

Ryan Sciampacone – IBM Managed Runtime Architect 14th June 2013

© 2013 IBM Corporation 2

Important Disclaimers

THE INFORMATION CONTAINED IN THIS PRESENTATION IS PROVIDED FOR INFORMATIONAL PURPOSES ONLY.

WHILST EFFORTS WERE MADE TO VERIFY THE COMPLETENESS AND ACCURACY OF THE INFORMATION CONTAINED IN THIS PRESENTATION, IT IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED.

ALL PERFORMANCE DATA INCLUDED IN THIS PRESENTATION HAVE BEEN GATHERED IN A CONTROLLED ENVIRONMENT. YOUR OWN TEST RESULTS MAY VARY BASED ON HARDWARE, SOFTWARE OR INFRASTRUCTURE DIFFERENCES.

ALL DATA INCLUDED IN THIS PRESENTATION ARE MEANT TO BE USED ONLY AS A GUIDE.

IN ADDITION, THE INFORMATION CONTAINED IN THIS PRESENTATION IS BASED ON IBM’S CURRENT PRODUCT PLANS AND STRATEGY, WHICH ARE SUBJECT TO CHANGE BY IBM, WITHOUT NOTICE.

IBM AND ITS AFFILIATED COMPANIES SHALL NOT BE RESPONSIBLE FOR ANY DAMAGES ARISING OUT OF THE USE OF, OR OTHERWISE RELATED TO, THIS PRESENTATION OR ANY OTHER DOCUMENTATION.

NOTHING CONTAINED IN THIS PRESENTATION IS INTENDED TO, OR SHALL HAVE THE EFFECT OF:

- CREATING ANY WARRANT OR REPRESENTATION FROM IBM, ITS AFFILIATED COMPANIES OR ITS OR THEIR SUPPLIERS AND/OR LICENSORS


Introduction to the speaker

■  16 years experience developing and deploying Java SDKs

■  Recent work focus: ■  Managed Runtime Architecture

■  Java Virtual Machine improvements

■  Multi-tenancy technology

■  Native data access and heap density

■  Footprint and performance

■  Garbage Collection

■  Scalability and pause time reduction

■  Advanced GC technology

■  My contact information: – [email protected]


What should you get from this talk?

■  JVM proving to be a fertile ecosystem for languages

■  Plenty of opportunity to innovate in other spaces

■  Runtime is the gateway to this innovation

■  Largely ignored the last few years, but this is where the core inventions can occur


The runtime isn’t boring!


Packed Objects


Problem? What problem?

■  JNI just isn’t a great way to marshal data

■  Locality in Java can matter (e.g., JEP 142)

■  Existing native and data placement stories aren’t very good

■  In many cases, legacy systems exist – the interop is just terrible

■  So we want something that integrates well with the Java language and helps us…


What are we trying to solve?

Simple enough…

Hash

Array

Entry

(object)

(object)

Object header

Object field / data

table key

value



Simple enough…

■  Header overhead

Hash

Array

Entry

(object)

(object)

Object header

Object field / data

table key

value



Simple enough…


■  Pointer chasing

Hash

Array

Entry

(object)

(object)

Object header

Object field / data

table key

value



Simple enough…


■  Pointer chasing

■  Locality

Hash

Array

Entry

(object)

(object)

Object header

Object field / data

table key

value



int int

Object header

Fields / data

Java Native

…

int[] d

anObject

Fighting the Java/Native interface

int int

int …


Ok so we have some criteria…

■  Ability to do away with headers

■  Ability to bring multiple objects close together

■  On heap / off heap seamless referencing of data

■  This actually sounds a lot like C structure types

■  Packed Objects!

struct Address { char[4] addr; short port; } struct Header { struct Address src; struct Address dst; }

struct Header

port addr Address src

Address dst

port addr


Packed Objects: Under the covers

int y int x

aPoint

Object header

Object field / data



int y int x

aPoint

Object header

Object field / data

offset target

aPackedPoint

int y int x


Packed Objects: In Practice

int y int x

aPoint Object field / data

int y int x

aPoint Point e Point s

aLine

Object header



int y int x

aPoint

Object header

Object field / data

int y int x


aLine

aPackedLine

int y int x

int y int x

offset target



int y int x

aPoint

Object header

Object field / data

int y int x


aLine

int y int x

int y int x

aPackedPoint s

aPackedPoint e

aPackedLine

offset target



int y int x

aPoint

Object header

Object field / data

int y int x


aLine

int y int x

int y int x

aPackedPoint s

aPackedPoint e

@Packed final class PackedPoint extends PackedObject {

int x; int y;

} @Packed final class PackedLine extends PackedObject {

PackedPoint s; PackedPoint e;

}

aPackedLine

offset target



int y int x

aPoint

Object header

Object field / data

int y int x


aLine

int y int x

int y int x

aPackedLine

offset target



int y int x

aPoint

Object header

Object field / data

int y int x


aLine

int y int x

int y int x

aPackedLine.e

aPackedLine

offset target



int y int x

aPoint

Object header

Object field / data

int y int x


aLine

int y int x

int y int x

aPackedLine.e

offset target

aPackedPoint

aPackedLine

offset target


Packed Objects: In Practice with Native Access

int y int x

Object header

Struct field / data

Java Native

struct Point { int x; int y; }



int y int x

Object header

Struct field / data

offset target

aPackedPoint

Java Native



int x; int y;

}



int y int x

Object header

Struct field / data

offset target

aPackedPoint

Java Native


Ø


int x; int y;

}


Lets Build Something in C!

■  Nested substructures

■  Compact representation

■  Alignment aspects

struct Address { char[4] addr; short port; } struct Header { struct Address src; struct Address dst; }

struct Header


Address dst

port addr


Let’s Build the Same “Something” in Java!

■  Headers

■  No locality

■  Alignment

class Address { byte[] addr; short port;

} class Header {

Address src; Address dst;

}

Address

port addr addr Header

Address dst Address src

Address

port addr addr

byte[]

byte[]


What does the Java code look like under the covers?

■  From a code point of view, this isn’t terrible…

Bytecodes: aload1 getfield Header.dest LAddress; getfield Address.addr [B iconst0 baload bipush 192 ificmpeq ...

JIT (32 bit): mov EBX, dword ptr -4[ECX] // load temp1 mov EBX, dword ptr 8[EBX] // load dest mov EBX, dword ptr 4[EBX] // load addr movsx EDI, byte ptr 8[EBX] // array[0] cmp EDI, 192


What if we did this with Packed Objects?

■  The Java code is pretty clean… and a pretty good result!

@Packed final class Address extends PackedObject {

PackedByte[[4]] addr; short port;

} @Packed final class PacketHeader extends PackedObject {

Address src; Address dest;

}


Address dst

port addr

offset target

aPackedHeader


Ok, what about the code under the covers?

■  Bytecodes don’t change… JIT code is pretty good too!

JIT (32 bit): mov EBX, dword ptr -4[ECX] // load temp1 mov EAX, dword ptr 4[EBX] // load target mov EDX, dword ptr 8[EBX] // load offset lea EBX, dword ptr [EAX + EDX] movsx EDI, byte ptr 8[EBX] // array[0] cmp EDI, 192

Bytecodes: aload1 getfield PackedHeader.dest LAddress; getfield Address.addr [B iconst0 baload bipush 192 ificmpeq ...


What about native access?

Java Native

…

anObject

How do we implement this normally?

port addr

port addr


JNI implementation

■  Usual “stash pointers in long types” tricks

■  JNI costs tend to be high


DirectByteBuffer implementation

■  No extra JNI to write (this is good)

■  Keeping your indices straight is never fun


Unsafe implementation

■  You shouldn’t be here

■  Still playing the indices game


PackedObject answer

■  Looks like natural Java code

■  Foregoes JNI

■  Same type capable of on-heap representation

port addr

port addr offset

target

aPackedHeader

Ø


Active work and next steps

■  Experimenting with this now

■  Yes, there are security aspects to be aware of here

■  This is potentially part of a larger look at Java / Platform interop

■  Not specifically viewed as a cure to GC problems

■  This forms the basis for many other solutions to existing problems…


Multitenancy


Just what do you mean by “multitenancy”?

With a multitenant architecture, a software application is designed to virtually partition its data and configuration, and each client organization works with a customized virtual application instance.

■  Working Definition – A single instance of a software application that serves multiple customers

à Each customer is a tenant. – Tenants can customize some parts of the application (look and feel) but not the code. – Infrastructure usually opaque

à opportunity for provider

Why? Cost Savings: As compared to single-tenant deployment model


S5. Shared Application

JDK Support for Spectrum of Sharing / Multitenancy (Level 1-5)

S1. No Sharing

Infrastructure

Middleware

Application

Tenant

OS

Middleware

Application

Tenant

OS

Infrastructure

Middleware

Application

Tenant

OS

Middleware

Application

Tenant

OS

Infrastructure

Data Center floor Data Center floor

Middleware

Application

Tenant

Middleware

Application

Tenant

Infrastructure

Data Center floor

OS

Application

Tenant

Application

Tenant

Infrastructure

Data Center floor

OS

Middleware

Application data

Tenant Tenant

Infrastructure

Data Center floor

OS

Middleware

Application

S2. Shared Hardware

S3. Shared Operating System

S4. Shared Middleware

Application data

Application data

Application data

Application data

Application data

Application data

Application data

Application data

Application data

Sharing servers storage, networks in a data center

Hypervisors (e.g. KVM, VMWare) are used to virtualize the hardware

Multiple applications sharing the same middleware

Sharing the same application

Multiple copies of middleware in a single operating system

Application Changes

Application Changes ? Application

Changes


= L Application Changes


Hardware Virtualization

■  Hypervisors run multiple applications side-by-side safely

■  Advantages – Capture idle CPU cycles – Automatic de-duplication (RAM) – Ability to meter and shift resource toward demand – No need to change tenant applications

Hypervisor

Hardware

tenant tenant tenant tenant


Hardware Virtualization

■  Hypervisors JVMs can run multiple applications side-by-side safely

■  Advantages – Capture idle CPU cycles – Automatic de-duplication (ability to share Java artifacts) – Ability to meter and shift resource toward demand – No need to change tenant applications

Hypervsisor

Hardware

tenant tenant tenant tenant

Java VM

Operating System


Multitenancy: Low (or no) barrier to entry

■  Multitenancy is all about reducing duplication by transparently sharing a JVM – 1 GC, 1 JIT, shared heap objects – plus: JVM-enforced resource constraints to meter and limit consumption

■  Ergonomics: Opt-in to multitenancy with a single flag: -Xmt (multitenancy) – no application changes required

javad

Tenant1

Tenant2

One copy of common code + data lives in the javad process.


JVM: Separating State

■  Static variables are a problem for sharing

■  Consider use of System.out in code we want to share below



getstatic does 2 things 1.  Triggers class initialization on first contact

–  Danger: Each ‘tenant’ needs to do this 2.  Resolves a name (out) to a storage location and reads from it

–  Danger: Each tenant needs dedicated storage


JVM: Standard Static Fields

getstatic <cpIndex> 0: if constantPool[cpIndex] is resolved then { 1: fetch field address from constantPool 2: read field value from class 3: return field value 4: } else { 5: initialize class 6: determine address of field 7: store field address into constantPool 8: goto 1 9: }


JVM: Isolated Static Fields

getstatic <cpIndex> 0: if constantPool[cpIndex] is resolved then { 1: fetch field address index and class from constantPool 2: fetch tenant->data[class->initIndex] 3: if (!initialized) { 4: initialize class; 5: } 6: read field value from tenant->data[index] 7: return field value 8: } else { 9: initialize class 10: determine address index of field 11: store field address index into constantPool 12: goto 1 13: }



■  There are other initialization triggers too – putstatic – invokestatic – instanceof – new

■  Some of these warrant special handling as they are heavily used (invokestatic) and initialization checks are expensive.

– Separate stack frame build from opcode


More that just JVM State…

■  Throttling of resources – Threads, GC, sockets, files (IO in general), native memory

■  Past and existing examples do exist! – Commercial / In house custom solutions – JSR 181 Isolates / 284 Resource Management

■  Security is of course huge


Other Thoughts – Native Libraries and shared state

■  Use separate processes to manage different state

■  Each process now holds the context

■  Challenges: Latency

Tenant1

Tenant2

JVM

Proxy Library

Shared Library

Shared Library

Tenant1

Tenant2

JVM

Shared Library

Shared State!

■  Native libraries contain state that may not be shareable across tenants


Heap Memory (not Garbage Collection!)


Scaling Garbage Collection

■  Existing Garbage Collection (GC) technology defers expensive costs

Heap

Local Collection

Global Collection

Heap

■  Global approach introduces linear scaling problem (100ms @ 1GB → 10s @ 100GB)

■  Goal: Scale using result-based, incremental (through partitioned heap) garbage collection

■  Target appropriate data sets based on environment factors

■  Key: Region Based Garbage Collection approach


Challenge: Segregate Memory by Differentiators

Heap Varying memory

characteristics and requirements per

application

JVM must dynamically

recognize and adapt given platform

resources

IaaS optimization + sharing (consolidate) SSD Exploitation (memory efficiency) Multi-core scalability (scale up)

Very Large Heap GC (low pause times) Soft Real-Time

■  Objects with common properties •  Locality •  Usage frequency •  Lifetime •  Levels of “Read-only”ness

■  Optimize: Memory characteristics •  Page sharing status •  Memory speed aware •  Flash / SSD exploitation

■  “Results based” incremental operations

•  Productive GC every cycle •  Localized garbage collect

Custom

Placement improves cache performance Allocation efficiency (thoughtput) Reduced size working set (paging)


Memory Hierarchy and Performance

■  Increased core and HW thread counts are not silver bullets to scaling in a memory intensive environment such as Java

■  Pressure from increased cores / HW threads

■  Cache hierarchy does not scale Cache

Hierarchy

BUS

Vs.

Logical Processors

Memory

Memory

Memory

Memory

CPU CPU

CPU CPU

■  Memory speed varies based on execution context

■  Magnitude of variance related to CPU socket count

■  Uneven application performance over time •  Data access can be fast or slow (or both) •  Behavior can be variable run to run due to data placement •  Threads scheduling can be unpredictable •  Same deployments on different hardware cause unpredictable performance shifts


Let’s look at transferring data

Heap

A

B

C



Heap

A

B

C

Remote Transfer


PackedObjects could help…

Heap

A

Remote Transfer

C

B



Heap

A

B

C

Remote Transfer

A

Packed

C

B



Heap

A

B

C

Remote Transfer

A

Packed

B

C

A

Packed

C

B


Making the data transfer easier…

Heap

Remote Transfer



Heap

Remote Transfer

Specialized Heap Area



Heap

B

Remote Transfer

A B

C

A

C



Making the data transfer seamless

Heap

Remote Transfer



B

C

A

B

C

A


Heap

Remote Transfer

B

C

A



B

C

A

B

C

A


Heap

Remote Transfer

B

C

A Packed


offset target

Packed

offset target


Tip of the Iceberg

■  Just scratching the surface of what is possible

■  GC technology has an effect on what is possible – But this should be invisible to the application and seamless in the language

■  Innovation in the heap space is not limited to GC


Virtualization


Finding your way in The Cloud

■  Constant pressure to reduce cost and deliver more services for a given cost

■  Maximize the available resources (Hardware)

■  CPU and Memory are typically the biggest – Other critical ones as well, including I/O

■  Plenty of solutions – Cloud and Multitenancy just to name a few

■  And so by and large we’re in a Virtualization scenario


The Layers of Liars

■  Your application could end up querying 3 different layers to get a (hopefully) correct picture

Hypervisor

Hardware

OS

API! availableProcessors() [no accuracy / understanding]

Thinks it knows but doesn’t (being lied to)

Knows and controls, but can change its mind

App


The Layers of Liars

Hypervisor

Hardware

App

OS

App

Image source (Wikipedia Commons) CPU: https://en.wikipedia.org/wiki/Central_processing_unit DRAM: http://en.wikipedia.org/wiki/Dynamic_random-access_memory

RAM

CPU


The Layers of Liars

Hypervisor

Hardware

App

OS

App


RAM

CPU


The Layers of Liars

Hypervisor

Hardware

App

OS

App App

OS


RAM

CPU


The Layers of Liars

■  You don’t even have a constant picture of the universe

Hypervisor

Hardware

App

OS

App App

OS


RAM

CPU


Knowing what you don’t know

(What does this even mean anymore?)


Interfaces… more importantly behavior

■  Provide interfaces to the “Layers of liars” to enable middleware “smarts”

■  Runtime reads and reacts (in co-operation with application / other runtimes)

Hypervisor

Hardware

App

OS

Hypervisor

Hardware

App

OS

App App

OS

Adjust –Xmx Adjust –Xgcthreads Adjust thread pool counts …


…And after all that


Tenant

Nimble Reactions to External Events

Hypervisor

Hardware

OS

Application

Tenant


Tenant


Hypervisor

Hardware

OS

Application

Tenant


Tenant Tenant


Hypervisor

Hardware

OS

Application App

OS

Tenant


Tenant Tenant Tenant Tenant


Hypervisor

Hardware

OS

Application App

OS

Tenant

Hypervisor

Hardware

OS

Application


Tenant Tenant Tenant Tenant


Hypervisor

Hardware

OS

Application App

OS

Tenant

Hypervisor

Hardware

OS

Application

Tenant

RDMA + Packed Objects


Questions?


References

■  Get Products and Technologies: – IBM Java Runtimes and SDKs:

•  https://www.ibm.com/developerworks/java/jdk/ – IBM Monitoring and Diagnostic Tools for Java:

•  https://www.ibm.com/developerworks/java/jdk/tools/

■  Learn: – IBM Java InfoCenter:

•  http://publib.boulder.ibm.com/infocenter/java7sdk/v7r0/index.jsp

■  Discuss: – IBM Java Runtimes and SDKs Forum:

•  http://www.ibm.com/developerworks/forums/forum.jspa?forumID=367&start=0


Copyright and Trademarks

© IBM Corporation 2012. All Rights Reserved.

IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of International Business Machines Corp., and registered in many jurisdictions worldwide.

Other product and service names might be trademarks of IBM or other companies.

A current list of IBM trademarks is available on the Web – see the IBM “Copyright and trademark information” page at URL: www.ibm.com/legal/copytrade.shtml

runtime innovation - nextgen ninja hacking of the jvm, by ryan sciampacone

Technology