YOW! JVM Demystified Dave Thomas

1

•  Dates – Melbourne: December 5-6 – Brisbane: December 9-10 – Sydney: December 12-13

•  Early bird tickets on sale early June! •  Don’t forget YOW! 2012

presentations, videos, interviews and podcasts all available online

YOW! 2013

YOW! 2013 Ini-al Speakers

Jeff Hawkins Co-Founder of Palm, author of “On Intelligence”,

artificial intelligence & the brain

Confirmed Keynotes

2


Scott Hanselman Microsoft Principal Project Manager, author of

numerous ASP .NET books, host of Hanselminutes and other podcasts

Confirmed Keynotes


Philip Wadler Computer Scientist known for his contributions to

programming languages and type theory

Confirmed Keynotes

3


§ Kevlin Henney Developer Best Practices

§ Doug Crockford JavaScript guru, inventor of JSON

§ Daniel Spiewak Scala fanatic, functional devotee

§ Scott Murray D3.js

§ Mike Dixon Big Data, google.gov

§ Jesper Richter-Reichhelm Scalability of game servers at Wooga

§ Elizabeth Keogh Agile expert, human psychology

§ Aaron Bedra Application Security at Braintree

•  Ben Christensen RxJava, Netflix

§ Phillipe Kruchten Architecture, creator 4+1 model

§ Jim Webber NoSQL, Neo4J, REST

§ Michael Nygard DevOps, author of “Release It”

§ Steve Vinoski Erlang, IEEE Internet Computing

§ Trish Gee MongoDB, Java high performance

§ Eric Evans Domain Driven Design

§ Charles Nutter Jruby

§ Hadi Harri JetBrains, Kotlin

© 2013 Bedarra Research Labs. All rights reserved.

JVM Demys*fied – A Tour of the LVM Engine Room

Dave Thomas www.davethomas.net

¡ Bedarra Research Labs ¡ YOW! Conferences

¡ Carleton University ¡ Queensland University of Technology ¡

4

Outline

Hardware Machines

Virtual Machines

Essence of a Language Virtual Machine

700+ languages in 7 minutes

LVM Engineering Design Choices

LVM Speed and Space

Musings on Next Genera*on VMs

Discussion



Real Machines

5

Modern Hardware (Intel Sandy Bridge EP*)

C 1 C n C 1 C n Registers/Buffers <1ns

L1 L1 L1 L1 ~4 cycles ~1ns

L2 L2 L2 L2 ~12 cycles ~3ns

L3 L3 ~40 cycles ~15ns

~60 cycles ~20ns (dirty hit)

~65ns

DRAM

QPI ~40ns MC MC

DRAM

DRAM

DRAM

DRAM

DRAM

DRAM

DRAM

...

...

...

...

...

...

QPI QPI PCI-e 3 PCI-e 3

40X IO

40X IO

* Assumption: 3GHz Processor (c) Martin Thompson 2012 - 2013

6


Real Hardware a Moving Target

Real HW Virtual Machines (Virtualiza*on)

§  SoZware which seeks to efficiently emulate a specific real machine

§  VM360, VMWare, XEN, VirtualBox , Azure, Parallels …

§  Simplify deployment; isolate users; share resources


7

Real Language Machines

Hardware which seeks to directly support a high level language features -‐Tags, Write Barriers, Contexts ..

ALGOL Burroughs 6700, APL IBM 5100, Lisp Symbolics, XPARC Dorado, PICK, Prolog, ADA Intel 432, Smalltalk SOAR, Java Azul; …


Virtual HW Machines

§  SoZware which seeks to efficiently emulate an ideal abstract HW machine on different real HW plajorms

§  IBM AS/400, Calusa B-‐RISC, Google Portable Na*ve Run*me


8


Language Virtual Machines (LVM)

§  SoZware which seeks to efficiently implement a programming language by emula*ng a language specific machine representa*on such as byte codes

§  Pascal P-‐Code, SECD, LVMs for Smalltalk, Java, CLR, Perl, R, MalLab, Erlang, PHP, Python, Ruby, JavaScript, Dart …

Hardware Machines Language Virtual Machines

Instruc*on Set Abstract instruc*on set (AST, byte Codes…)

Registers and Memory Stack + Memory

Fetch – Decode – Execute Read – Eval – Print Loop (REPL)

Storage Manager (GC)


Why have LVMS? – Just Compile It!

•  Compilers and Interpreters Easy

•  Security

•  Execu*on Time Op*miza*on


9


400+ Languages in 4 minutes



The READ – EVAL – PRINT LOOP aka “REPL”

Meta circular Interpreters and the meaning of Life

The Contour Model of Block Structure Processes

10



The READ – EVAL – PRINT LOOP aka “REPL”

What are the key execu*on *me differences between these languages?


VM Engineering Design Space

11


VM Instruc*on Sets

Essen*al or Exo*c?


12


Boxing and Tags

Boxed Languages

§  Na*ve machine types

§  No fully polymorphic collec*ons without boxing

Tagged Languages

§  Fully Polymorphic collec*ons

§  Tag bits take space and need checking (SOAR => Sparc)

VM Implementa*on

Interpreter, Compiler, Mixed, Mul*ple …

Low Level Language – C/C++ , C

Implement the LVM in a subset of the LVM language (self hos*ng) •  Use an exis*ng implementa*on of language or a similar one •  Use a your new language and manually translate implementa*on to low

level •  Implement the reflec*ve tower e.g. 3-‐Lisp

Base Types -‐ implemented inside the LVM e.g. Integer, Object…

§  How easy should it be for others? §  Opera*ons, New types (CLR Value Types), New representa*ons §  Access to HW

Core Libraries •  Inside LVM or Open in Language?

©2003 Bedarra Research Labs. All rights reserved.

13


Implementa*on -‐ Core Libraries

Implement in the language or in the LVM?

Library in Language Implement in LVM


Method/Func*on Dispatch

§  Tables, Dynamic Lookup/ Dynamic Transla*on

§  Polymorphic Inline Cache (PIC)

§  Run*me Type Feedback (Self-‐> Strongtalk-‐> Hotspot)

§  Profile Guided Inlining (also called tracing JIT)

§  Whole Program Compila*on – Smart Eiffel, Ahead-‐Of-‐Time (AOT) JIT IBM, Excelsior Jet …

14


Garbage Collec*on


Concurrency Model

•  Actors versus Data Parallel §  Immutable Data Structures

§  Low level Threads and Locks (semaphores)

Na*ve OS Threads

versus Green Threads

15


Applica*on Performance versus Benchmark Lies

Very Few Know Where the Time and Space Really Goes!

Micro benchmarks such as SPEC, Sun Spider, Octane

§  Measure specific language features

§  Seldom a predictor for applica*on performance

1000000 timesRepeat[12 fibonacci]

Applica*on Performance

§  Dominated by macro behavior of applica*on

§  Time spent in OS or DB

§  Time spent in Framework/Library

§  Applica*on or Framework races GC

§  Applica*on or Framework thrashes cache


Your Mileage will Vary!

Client Centric Applica*on §  Ample but limited memory 4 – 8G

§  Ample but limited CPU Server Centric Applica*on §  Significant GUI overhead Space per instance

§  Startup time per instance §  Persistent Object/Record Serialization and Materialization §  Ample memory 64-1TB + cpu 32 – 128 cores

Embedded or Pervasive §  Power Constrained and Instant On

§  Memory Constrained

§  Code in Flash

§  Trusted

§  Na*ve Code Interoperability

§  Real-‐*me GC

16

Java JVM Challenges Today

Java Developers s*ll face big challenges §  “Cloud” -‐ virtualiza*on, footprint/density, run*me dynamism

§  “Big*.*” -‐ data, threads, memory, scale out, scale up

§  Security -‐ cri*cal to connected soZware and hard to get right §  Compa*bility – need to innovate and protect exis*ng investment in

code and tools

§  Plus ... mobile , polyglot , development efficiency, simplicity

© John Duwimovich CTO IBM Java VMs “ We can do beler …”



Challenges with JVM and CLR… •  Stacks easy for compiler, hard for JIT •  Security difficult at best and limits access to some na*ve HW

features •  Difficult to exploit major changes in HW e.g. memory models,

massive concurrency… •  LVM are always L Biased to the LVM family of languages

§  E.g. strong types, no tags, no call cc…

Is it possible to have a UVM which lets interesOng languages bloom and prosper to avoid bloaOng My Language PL/I, Ada,

Java 7+, C++ 11 …?


17

Further Reading

Structure and Interpreta*on of Computer Programs, Abelson and Sussman

The Lambda Papers, Guy Steele, MIT Technical Reports

Understanding Java Garbage Collec*on, Gil Tene

JVM JIT for Dummies, Charles Nuler

We Can Do Beler, IBM’s Vision for Next Genera*on Java VMs, EclipseCon Europe

Smalltalk-‐80: The Language and its Implementa*on, Goldberg, A., Robson. I, Addison-‐Wesley, 1983.

Op*mizing Dynamically-‐Dispatched with Run-‐Time Type Feedback,

Urs Holzle

V8 A High Performance Open Source JavaScript Engine, Kasper Lund

Trace-‐based Just-‐in-‐Time Type Specializa*on for Dynamic Languages, Andreas Gal et al



Back to the Future? – Virtual Instruc*on Set Computers (VISC)

•  AS/400, DAISY, B-‐RISC, Transmeta, LLDA, AMD HAS, Google PNaCl…

•  Supports many language using so7ware fault isola-on •  Infinite register machine •  Single assignment 3 address instruc*ons •  Op*onal data flow and control flow informa*on


18



•  JIT Virtual Machine to Real Machine •  LLVM low level infrastructure popular plajorm •  Google PNaCl “Pinnacle” claim 7% overhead on Intel and 5% on

ARM for C/C++ •  AMD Heterogeneous System Architecture (HAS) VSIC for CPU+

GPU …

Thanks! Please give them the applause


YOW! JVM Demystified Dave Thomas

Documents