PennBench: A Benchmark Suite for Embedded Java

Transcription

1 WWC5 Austin, TX. Nov PennBench: A Benchmark Suite for Embedded Java G. Chen, M. Kandemir, N. Vijaykrishnan, And M. J. Irwin Penn State University

2 Outline Introduction to Embedded Java Introduction to PennBench Suite Characterizations Method invocation behavior Memory allocation behavior Conclusions

3 Introduction to Embedded Java

4 J2ME Java for Embedded Devices From SUN Microsystem

5 Why Java? 1. Dynamic delivery of applications and services 2. Cross-platform compatibility 3. Enhanced user experience 4. Disconnected access 5. Security We can allow third-party application developers who we don't know to put applications in the telephone because we know that Java TM technology will protect the phone." -- Brian Modra, software architect, Nokia

6 Architecture of a Java Enabled Device MIDP Applications OEM Spec Applications OEM Spec API MID Profile CLDC/KVM Device OS

7 MIDP (Mobile Information Device Profile) & CLDC (Connected Limited Device Configuration) Standardized Java APIs Developed utilizing Java Community Process (JCP) Program Supported by major mobile phone manufactures Targeting at resource constrained devices Limited memory: usually no more than 1MB Limited CPU speed Limited energy supply: powered by batteries Ex. mobile phones, entry level PDAs

8 PennBench Suite

9 Why not SPEC JVM98 Benchmarks? Heap Memory Computation MIDP Devices Typically <= 1MB Usually Light 10MB to run, 32MB+ for reasonable performance Heavy SPEC JVM98 User Interactivity Networking Typically Yes No No

10 Guidelines for benchmark selection Be able to run on MIDP/CLDC devices Include both utility and entertainment applications Include different types of applications CPU intensive, Memory intensive, and GUI intensive Touch all the major components of a MIDP/CLDC device Be useful or interesting Publicly available

11 Method Invocation Behavior of PennBench Suite

12 Method Invocations Breakdown User Library Invocations Dynamic Bytecodes 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% calculator chess viewer jpegviewer manyballs mdoom mpg mvideo pushpuzzle scheduler sfmap webviewer calculator chess viewer jpegviewer manyballs mdoom mpg mvideo pushpuzzle scheduler sfmap webviewer

13 Just-In-Time Compilation and Hot Methods Bytecode interpreter is slow JIT - dynamically compile the bytecode into native code at runtime JIT has heavy overheads The compilation cost is amortized in the future invocations of the compiled methods Only the hot (frequently used) methods can pay off the compilation cost How many methods are hot? How to identify hot methods?

14 Hot methods Accumulated Number of Invocations 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% calculator chess viewer jpegviewer 0% 5% 10% 15% 20% 25% Number of Methods

15 Threshold for Just-In-Time Compilation 60% calculator chess 50% viewer jpegviewer Compiled Methods 40% 30% 20% 10% 0% Threshod: Number of Invocation

16 Distribution of Method Sizes 14% 12% 10% calculator chess viewer jpegviewer 8% 6% 4% 2% 0% Size (Bytes)

17 Memory Allocation Behavior of PennBench Suite

18 Accumulated Memory Footprints Accumulated Size (KB) Live Objects Allocated Dead Time (seconds)

19 Memory Footprints (Cont.) Live Object Size (KB) Minimal Heap Size Time (seconds)

20 Generational Garbage Collection Most objects live for very shot time A small percentage of them live much longer Free Young Generation Old Generation Free Young Generation Old Generation Old Generation Free Young Generation Old Generation

21 Distribution of Object Lifetime 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% calculator chess viewer jpegviewer manyballs mdoom mpg mvideo pushpuzzle scheduler sfmap webviewer 256B 1KB 4KB 16KB Other Permanent

22 Method Local Object Object O is local to method M if: O is created in method M; and O becomes garbage before M returns Stack Allocation Allocate method local objects in the stack frame of their creating method Method local objects are automatically collected when the frames are popped up due to method return or uncaught exceptions

23 Thread Local Object Object O is local to thread T if: O is created by thread T; and No thread other than T accesses O A method local object is also thread local Elimination of synchronizations Scheduler-controlled memory turnoff: Allocate thread local objects in the memory banks private to the thread The scheduler sets the memory banks of non-active threads to sleep mode to conserve energy

24 Method and Thread Local Objects 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% calculator chess viewer jpegviewer manyballs mdoom mpg mvideo pushpuzzle scheduler sfmap webviewer Method Local Thread Local Nonlocal

25 Conclusions Our work confirms the observations that have been made on high performance Java systems Only a small portion of methods are hot Generational behavior of objects A big portion of objects are method/thread local

26 Conclusions (cont.) A benchmark suite for embedded Java is necessary Most benchmarks for high performance systems are not able to run in embedded environments The system configuration parameters that are optimal for resource-rich systems may not be optimal in resource-constraint environments

27 Thank you!

28 PennBench Suite Benchmark calculator chess viewer jpegviewer Description A calculator running on MIDP Play chess game with computer A light-weight implementation of POP3 client Display a JPEG image on the screen

29 PennBench Suite (cont.) Benchmark manyballs mdoom mpg mvideo Description Draw many bouncing balls on the screen; each ball corresponds to a Java thread 3D shooting game MPEG movie player (local) Stream movie player (through HTTP)

30 PennBench Suite (cont.) Benchmark pushpuzzle Description A conventional puzzle game scheduler Personal monthly scheduler sfmap Digital map on MIDP webviewer WWW browser

31 Event Play Back Mechanism Application KVM Event Dispatcher Event File OS Keyboard Touch Screen Timer

32 Property access method Java code Bytecode class String { private int count; int length() { return count; } } Method int length() 0 aload_0 1 getfield #22 4 ireturn