Remote ProgrammingDisseminationCollectionNetwork Management

Gilman Tolle(also speaking for Jonathan Hui)

A Sensor Network Link?

A Sensor Network Link!

Why Over-The-Air Reprogramming?

Embedded nature of sensor networks – they’re small! Network scales reaching thousands of nodes – there’s a lot of them! A necessity in debugging and testing cycle – we can’t stop messing! Learn about the environment after deployment – things change!

sensing data, network characteristics, etc.

What is Deluge?

A reliable data dissemination protocol for program images over a multihop network.

Combined with a bootloader (TOSBoot) Network Programming

0101010101010101011010101101010101010101

Program

Deluge Data Representation

Program divided into pages, each consisting of N packets.

Reduced RAM requirements Allows for spatial multiplexing

101110

110010

010000

101111

000011

Program

Packets

1234 N

How Does Deluge Work?

Nodes periodically advertise Suppress similar advertisements

Version 2 here.

I only have version 1.

I only have version 1.

How Does Deluge Work?

Neighboring nodes request data Suppress similar requests

Send me page 1!

Send me page 1!

How Does Deluge Work?

Requested data is broadcast

Packet 12of page 1!

How Does Deluge Work?

Dropped packets are NACKed

Repeat packet 4 of page 1!

Repeat packet 32 of page 1!

How Does Deluge Work?

Dropped packets are sent again

Packet 4 of page 1!

How Does Deluge Work?

Advertise for propagation to next hop

Version 2 here.

I only have version 1.

Spatial Multiplexing

Propagate in “waves”

Exploit limited radio range for concurrent broadcasts.

Reduced completion time o(d + Sobj) vs. o(d * Sobj)

Page 0Page 1

Epidemic Propagation

Epidemic propagation from one source

Deluge Features

Epidemic propagation from one source or many Continuous propagation effort by all nodes Turn on/off radios at will Reach nodes with intermittent connectivity Will find a path if it exists

Aggressive message suppression Scales with density Ultra low quiescent traffic

Deluge Features

Management Multiple program images Image metadata User confirmation on expensive operations

Minimize operator error

Robustness Redundant CRCs Golden Image with write protect Load Golden Image

Watchdog trigger Golden gesture

TOSBoot TOSBoot as isolated code Verify CRCs Verify system voltage

0101010101010101011010101101010101010101

0101010101010101011010101101010101010101

0101010101010101011010101101010101010101

ProgA ProgB ProgC

Program NameCompile TimeUserIDHostnamePlatform

CRASH!

CR

CC

RC

CR

CC

RC

CR

CC

RC

CR

CC

RC

CR

CC

RC

CR

CC

RC01010

10101010101011010101101010101010101

Deluge Lessons

Advantages Ease of reprogramming 100’s-1000’s of nodes Does not erase node IDs Golden Image is immensely useful Quickly switch between images More reliable than uisp or msp430-bsl Deluge over 802.15.4 more efficient that 802.11!

Disadvantages Ease of reprogramming 100’s-1000’s of nodes :)

Routing

Getting the packets through, across many hops Every node is a router too

Gateway-centric Get data from the gateway to all the nodes – dissemination Get data from all the nodes to the gateway – collection Node-to-node?

Message-based Transferring single packets, datagram-style

A layer, with clients above and services below Attribute Queries and Changes Event Reporting RPC Command Layer

IP is not the right solution Any-to-any is not useful enough to justify the state and complexity

Drip Dissemination Layer

Each sent message reaches all nodes in the network Good for sending commands and queries

A generic single-message communication layer command send[type](message) on the host event receive[type](message) on the node

Lightweight header – type and sequence number Higher layer can add destination addressing

Uses Trickle epidemic algorithm (Phil Levis) Dynamic forwarder selection Periodic retransmissions Neighborhood suppression

Caches latest message on each channel

GW

Client

Drain Collection Layer

Every node needs to send data to a gateway A generic single-message communication layer

command send[type](message) on the node event receive[type](message) on the host

Very well-studied problem (Too many authors to list) Link estimation plus distributed execution of shortest-path algorithm

Ours must support multiple gateways Each gateway builds a tree -- each node selects the cheapest next-hop Automatic subdivision of the network into pieces

GWGW

Client

Routing Goals

No predefined geographic structure Routing decisions based only on connectivity and link estimators Makes it easier to deploy and move nodes Minimal state – single next hop, update-in-place

Robust to lossy networks Drip periodically retransmits cached data with exponential backoff Drain uses link-layer ACKs, retransmissions, and long retry window Experiment with simple duty cycle: 1 second on, 1 second off Looks just like a lossy network, but the protocols keep working

Include rich metadata (source addresse, ttl, sequence numbers) “How are those bad packets getting into the network?”

Impossible to answer without metadata Enables network management through packet sniffing

Even without a specific “management” layer

Network Management

Sensor networks fail All networks fail sometimes!

Management lets us detect and respond to problems Just as important for our networks Harder to do, thanks to highly dynamic networks!

Passive Packet Sniffing

Active Network-Layer Monitoring

Management Queries

Monitoring Policies and Statistics

The Unbearable Lightness of Sniffing

Just sniffing packets reveals a wealth of information Active nodes from routed Drain messages Network topology from Drain beacon messages Dissemination behavior from Drip messages Reprogramming status from Deluge advertisement messages Overall traffic rates and histograms by type

Doesn’t make any extra demands on the network “If you’re going to send the packet anyway…” Passive management information gathering

Works when the network isn’t running the management layer Sometimes, you really need that extra few kB of code space

GW

The Story About Ping

Characterizing the performance of a dynamic network is HARD The first tool of network management: ping

Best with a few nodes that you don’t check very often The second tool of network management: the ping daemon

Can provide historical and current data on many nodes

The third tool of network management: the ping visualizer Scalable way to handle large complex networks Most networks don’t have a natural spatial realization

Our network is firmly embedded in space

The 10,000-Foot View

Management: The Gathering

Sometimes, you need to know more than ping can tell you How long has the node been running? Has the node been dropping any packets?

… and in our energy-constrained mote-land What is the node’s power source? Does the node have enough energy to run?

Management can be seen as a database problem TinyOS application exports information

Named attributes, variables in RAM Attribute schema generated at compile time

Replace long names with short integers – save bandwidth Store schema on mote using Deluge Supplement

Nucleus component responds to queries “Get attributes {2, 3, 6} and RAM variable at {0x1a8c} with length {2}”

System & App Components

Attribute Dispatcher

Query Processor

Drip/Drain

Nucleus Host Architecture

GW

Query Translator

Monitoring Daemon

Web Visualization

Command-Line Tools

GW

Packet Forwarder

Other Client Applications

XML-RPC

XML-RPC

Binary Data

XML-RPC

Binary Data

Pixels

Characters

Other Mgmt Tools