Mat

1
Maté A Tiny Virtual Machine for Sensor Networks

• Philip Levis and David Culler
• Presented by Michele Romano

2
Outline

• Sensor Networks
• Virtual Machines
• Maté Details
• Evaluation
• Conclusion

3
Sensor Networks

• Composed of 1000s of tiny devices (Motes) with
  limited resources

4
Berkeley Mote Specifications
5
TinyOS

• OS designed for sensor networks
• Split-phase non-blocking execution
• Not suited well to non-expert programmers

6
Reprogramming Motes

• Reprogramming is desirable as
• Environmental conditions change
• Analysis techniques evolve
• Examples
• Great Duck Island
• Building instrumentation

7
Reprogramming Motes

• To change the behaviour of a TinyOS program,
  either
• 1. Hardcode a state transition
• OR
• 2. Modify source code, recompile a TinyOS image
  and place image on mote

8
Sensor Networks Challenges

• Energy
• Recharging is difficult or impossible
• Deterministic network lifetime desirable
• Communication
• Lossy wireless networks
• Bandwidth conservation
• Programming
• Motes unreachable in deployed networks
• Difficult for a non-programmer to program TinyOS

9
System Requirements

• Small
• Expressive
• Concise
• Resilient
• Efficient
• Tailorable
• Simple

10
Virtual Machine

• Easier programming
• Short VM programs
• A VM can provide a safe program execution
  environment

11
Maté VM Overview

• Bytecode interpreter that runs on TinyOS
• Single TinyOS component that sits on top of
  several system components
• Code fits in capsules of 24 instructions
• Built-in ad-hoc routing algorithm AND mechanisms
  for writing new ones

12
Code Capsules

• There are four types of capsules
• Message send capsules
• Message receive capsules
• Timer capsules
• Subroutine capsules

13
Maté Architecture
14
Instruction Set

• There are three classes of Maté instructions
• 8 instructions reserved for users to define

basic 00iiiiii i instruction
s-class 01iiixxx i instruction, xargument
x-class 1ixxxxxx i instruction, xargument
15
Code Execution

• Execution of code begins in response to an event
• These three contexts can run concurrently
• Each instruction is executed as a TinyOS task

16
Code Security

• Bound checks prevent overrun and underrun
• Heap addressing is not a problem because there is
  only a single shared variable
• Unrecognized instructions result in no-ops

17
Code Infection

• Reprogramming is easy
• Each capsule contains a type and version number
• When a capsule with a more recent version is
  received, it is installed
• forw or forwo is used to broadcast the capsule
  for network neighbours to install

18
Maté Evaluation

• Ad-hoc routing algorithm was implemented to
  measure
• Rate of instruction
• CPU overhead
• Network infection rates

19
1. Rate of InstructionTest

• Maté Bytecode vs. Native Code

Operation Maté Clock Cycles Native Clock Cycles Cost
Simple and 469 14 33.51
Downcall rand 435 45 9.51
Quick Split sense 1342 396 3.41
Long Split sendr 68520000 20000 1.031
20
2. CPU Overhead

• Given the energy cost of an execution and the
  energy cost of installation
• Mate is preferable for a small number of
  executions
• For large number of executions, Native code is
  preferable

21
3. Network Infection

• Percentage of Motes Running New Program Over Time

22
Case Study

• Great Duck Island Application
• Spends most of its time in deep sleep mode
  draws 50 µA
• Reads several sensors and sends a packet
• Maté proves to save energy if only run for 5 days
  or less

23
Conclusion

• Maté met all of the defined requirements
• Maté can conserve energy in domains of frequent
  reprogramming
• VM can provide user-land guarantees

24
References

• http//www.cs.berkeley.edu/pal/research/mate.html
  
• http//www.cs.berkeley.edu/pal/pubs/brown-7-02.pd
  f
• http//www.cs.virginia.edu/qc9b/fall03cs851/mate_
  damon_jo.ppt