Darjeeling, a Feature-Rich VM for the Resource Poor

Niels Brouwers, Peter Corke, and Koen Langendoen

Introduction

Resource

Constrained

Modular,

Concurrent

WSN

Robust

Heterog

eneo

us

11/11/2009 2

Introduction

• Operating systems

• Contiki, TinyOS

• Native code (C/NesC)

• Efficient concurrency

Resource

Constrained

Modular,

Concurrent

• Efficient concurrency

• Virtual Machine

• Interpreted (Java)

• Robust, feature-rich

• Time/space overhead

WSN

Robust

Heterog

eneo

us

11/11/2009 3

Introduction

• We would like a VM that

• Enables a high-level language (Java)

• With an acceptable overhead

• Memory efficiency highest priority

• RAM

• Program flash

11/11/2009 4

Previous work

Feature rich Resource efficient Open source

Sentilla yes yes no

TinyVM (LeJOS) no yes yes

NanoVM no yes yes

Taka Tuka yes yes no

11/11/2009 5

Taka Tuka yes yes no

VM* yes yes no

MoteRunner yes yes no

Java Card no yes no

Squawk (Sun Spot) yes no yes

Darjeeling

• Feature rich

• Threading, synchronisation

• Compacting garbage collection

• Resource efficient

• Can run on devices with 2-10k RAM, ~64k flash

• But, no support for 64-bit or floating point, reflection

• Portable

• Contiki, TinyOS, FOS

• Open source

11/11/2009 6

Infuser tool

• Performs static linking

• Byte code post-processing

• Reduces code footprint by removing entity names

11/11/2009 7

Linked stack architecture

• Light-weight threads

• Stack frames are heap objects

• On demand stack space allocation

Stack Frame

Stack Frame

allocation

• Min memory usage: ~45 bytes Stack Frame

Thread

11/11/2009 8

16-bit architecture

• Custom instructions for 16-bit data types

• Saves stack space (operand stack, local variables)

• Arithmetic optimisation

11/11/2009 9

Double-ended stack

• Precise, compacting garbage collector

• Compaction requires identifying references on the operand stack

• Darjeeling uses two logical stacks

11/11/2009 10

Performance: Java vs C

30.4

77.781.5

54.47

113.2

40

60

80

100

120

Java/C30.4

0

20

40

MD5 Bubble sort

(32-bit)

Bubble sort

(16-bit)

8x8 vector

convolution

(32-bit)

8x8 vector

convolution

(16-bit)

Java/C

11/11/2009 11

Memory consumption: Stack Space

1819.4

21.7

15.114 14.3

17.3

12.115

20

25

32-bit

0

5

10

Blink CTP Test suite Base library

16-bit

11/11/2009 12

Springbrook

• Rewrite of a real-world application

• CTP implementation in Java

• 22 classes, ~1200 sloc

• <8kB in program flash• <8kB in program flash

• <2kB RAM

• 5 threads

11/11/2009 13

Springbrook power consumption

transmission

transmission

• CPU on-timeincreased by 50%

Time

PowerConsumption

cput

cpu

t

11/11/2009 14

Supported platforms

Tnode Tmote Sky Fleck3/Fleck3B

MCU Atmega128 MSP430 Atmega128(1)

Architecture 8-bit 16-bit 8-bit

RAM 4kB 10kB 4/8kBRAM 4kB 10kB 4/8kB

Radio CC1000 CC2420 nRF905

OS TinyOS Contiki FOS

Concurrency events protothreads threads

11/11/2009 15

Conclusions

• Darjeeling is an embedded VM

• Feature rich

• Memory efficient

• Open source• Open source

• Suitable for real-world applications

• Performance overhead acceptable for many applications

11/11/2009 16

Future work

• CLDC 1.0 compliance

• Support for the long data type

• Reflection

• Over-the-air loading• Over-the-air loading

• Generational garbage collector ?

• Operating system independent ?

http://darjeeling.sf.net

11/11/2009 17

Questions?

11/11/2009 18

Code size reduction

27,349

42,834

19,02320000

25000

30000

35000

40000

45000

Java .class

856

13,795

146

7,889

19,023

3,591

0

5000

10000

15000

20000

Blink (-83%) CTP (-71%) Test Suite (-55%) Base library (-75%)

Darjeeling .di

11/11/2009 19

Micro benchmarks: performance

Test C Java Java/C Instr./s

MD5 13.1s 399.7s 30.4 52,294

Bubble sort (32-bit) 0.3s 23.3s 77.7 60,618

Bubble sort (16-bit) 0.2s 19.3 81.5 71,526

8x8 vector convolution (32-bit) 9.1s 496.7s 54.47 65,454

8x8 vector convolution (16-bit) 4.6s 520.9s 113.2 71,512

11/11/2009 20

Micro benchmarks: code size

jar di reduction

Blink 856 146 83%Blink 856 146 83%

CTP 27,349 7,889 71%

Test Suite 42,834 19,023 55%

Base library 13,795 3,591 75%

11/11/2009 21

Micro benchmarks: stack size (1)

JVM DVM reduction

Blink 18 14 22%Blink 18 14 22%

CTP 19.4 14.3 27%

Test suite 21.7 17.3 20%

Base library 15.1 12.1 20%

11/11/2009 22

Micro benchmarks: stack size (2)

JVM DVM reduction

Blink 10 6 40%Blink 10 6 40%

CTP 11.4 6.3 45%

Test suite 13.7 9.3 32%

Base library 7.1 4.1 42%

11/11/2009 23

Portability

Tnode Tmote Sky Fleck3/Fleck3B

MCU Atmega128 MSP430 Atmega128(1)

Architecture 8-bit 16-bit 8-bitArchitecture 8-bit 16-bit 8-bit

RAM 4kB 10kB 4/8kB

Radio CC1000 CC2420 nRF905

OS TinyOS Contiki FOS

Concurrency events protothreads threads

11/11/2009 24

Portability: code complexity

Tnode Tmote Sky Fleck3/Fleck3B

Main 225 359 86

Radio 69 92 69

Sensors 184 157 110

Total 478 608 265

11/11/2009 25

Linked stack architecture

• Test runs three threads, started at the same time

• Each thread performs a binary tree test three times, with a 10 millisecond sleep in between

• Let the heap manager log the total stack size of each • Let the heap manager log the total stack size of each thread as the program executes

• Simulates periodic work, common in WSN

11/11/2009 26

Linked stack architecture

Peak memory

consumption

at 1335 bytes.

11/11/2009 27

Linked stack architecture

• Synchronize threads

• Only one thread may perform work at a time

• Work is serialised automatically

11/11/2009 28

Linked stack architecture

Peak memory

consumption

down to 535

bytes.

11/11/2009 29

Memory consumption: Stack Space

9.3

4.1

11.4

13.7

7.1

Test suite

Base library

16-bit

6

6.3

10

11.4

0 2 4 6 8 10 12 14 16

Blink

CTP 32-bit

11/11/2009 30