Practical Aggregate Programming with Protelis @ SASO2017

images/logo
Practical Aggregate Programming with Protelis
Danilo Pianini 1 Mirko Viroli 1 Jacob Beal 2
{danilo.pianini, mirko.viroli}@unibo.it
jakebeal@ieee.org
1Alma Mater Studiorum—Universit`a di Bologna
2Raytheon BBN Technologies
SASO Tutorial
2017-09-22 – Tucson, AZ, USA
Pianini, Viroli, Beal (UniBo, BBN) Programming with Protelis 2016-09-22 – SASO Tutorial 1 / 93

images/logo
Context
The IoT is becoming a crowded and complex place
Future and emerging Internet-of-things are witnessing:
devices increasingly wearable, mobile, embedded, ﬂying...
increasing availability of heterogeneous wireless connectivity
increasing availability of computational resources (device/edge/cloud)
increasing production and analysis of data, everywhere, anytime
⇒ business, security, and privacy concerns will probably be drivers as well

images/logo
Context
Programming models for mobile and IoT applications
Client side
Single-device program: objects, functions, actors, futures, tasks,
activities
Local intelligence
Client-server communication
SOA (Http/Rest), MoM (RabbitMQ)
Server side
Storage: OO, relational, NoSQL — in memory or on-disk
Coordination (orchestration, mediation, rules enactment)
Situation recognition (online/offline, mining, business intelligence,
stream processing)
Scalability in the server calls for cloud-ification
Not really orthogonal to the whole programming model
It often dramatically affects system design

images/logo
Context
Heterogeneity of deployment contexts

images/logo
Context
Implications
Where does programming eﬀort end up into?
Programs of clients and servers highly depend on:
Platform and API
Communication technology
Number, type, and dislocation of devices
⇒ Design and deployments hardly tolerate changes
⇒ IoT systems tend to be rigid, and costly to debug, evolve, and maintain
The technological result
Few of the opportunities of large-scale IoT are currently exploited
Virtually, any computational mechanism (sensing, actuation, processing,
storage) could involve spontaneous, adaptive cooperation of large sets of
devices
How many large-scale deployments of adaptive IoT systems around?
Where are the “Collective Adaptive Systems”?

images/logo
Outline
1 Introduction
Aggregate programming: idea and motivation
Protelis as practical AP language and workbench setup
2 Core language constructs
The boring stuﬀ: functions, operators, and the like
Stateful computation with rep
Sharing information with nbr
Domain restriction with if
3 Libraries
Frequently used patterns: towards reusable building blocks
Filling the abstraction gap: Protelis-lang
4 More about the toolchain
Making nodes move
Conﬁguring batches and exporting data
Using real world maps
Getting from simulations to devices

images/logo
Introduction
Outline
1 Introduction
3 Libraries
Making nodes move

images/logo
Introduction
A programming model perspective
What do we lack in large-scale IoT systems?
The plain old platform-independent programming abstraction
Delegating to the underlying platform everything but the core logic:
non-functional aspects (network performance, battery saving...)
deployment issues
The challenge
Let’s tackle the worst possible scenario:
Countless heterogeneous, moving devices unpredictably situated
Pervasive sensing and actuation
Contextual (spatio-temporal) behavior
Abstract away the infrastructure
We can’t rely on the cloud or fog
But we must be ready to exploit the opportunity they provide

images/logo
Introduction
Towards aggregate computing I
Desiderata
Abstractions well suited to the scenario
Techniques that promote solid engineering principles
Reusability
Composability
Modularity
A toolchain for designing, programming, and testing
As few constructs as possible, tractable semantics

images/logo
Introduction
Towards aggregate computing II
Aggregate computing
Based on the “field calculus”: a core, tractable calculus
Relies on the idea of “computational field”
Similar to functional languages
Promotes programming as “describing the problem” rather than
“hacking a solution”
Practical languages based on such calculus are available...
...as well as simulation environments where to perform testing
Somehow very different from languages rooted on other paradigms

images/logo
Introduction Aggregate programming: idea and motivation
Manifesto of aggregate computing
Motto: program the aggregate, not individual devices!
1 The reference computing machine
⇒ an aggregate of devices as a single entity fading to the actual space
2 The reference elaboration process
⇒ atomic manipulation of a collective data structure
3 The actual networked computation
⇒ a proximity-based self-org system hidden “under-the-hood”
neighborhood
device

images/logo
Computing with ﬁelds
Field: a distributed data structure mapping each device to a value
()0
1
+
-
1
-1
ef(0,1)
ef

images/logo
A look under the hood
The computational round semantics
Every device operates as follows:
1 Sleeps from some time
2 Wakes up, discards the old information, processes the new
information it may have received, and computes its program
3 Sends the results away, and goes back to sleep
Synchronization
No synchronization required
Devices should compute rounds at a similar frequency
The impact of very diverse frequencies depends on the speciﬁc
program

images/logo
Introduction Protelis as practical AP language and workbench setup
Protelis [PVB15]: aggregate programming in practice
The Field Calculus is a “theoretical object”
It can (and is) used for mathematical proofs, but provides no compiler
or runtime
Protelis is an attempt at making the higher order ﬁeld calculus
practically usable through:
An executable interpreter
A concrete C-family syntax
Interoperability with the Java Runtime
With access to the existing plethora of Java libraries
A set of tools for development, including Eclipse plugins

images/logo
Running Protelis programs
Protelis is to be used to program aggregate of devices
We need to run the programs on a network of devices
A single device is a very special case of a network of devices...
...but a purely local run does not normally provide any useful
information about the algorithm
We need to run the programs on a network of devices
Needless to say how impractical is to deploy a network every time you
want to perform a run
Another tool is required!
Protelis exposes an API for writing back-ends in charge of device,
hardware, and network interaction
It’s rather easy to make Protelis work upon existing frameworks...
...hereby including simulated environments

images/logo
Alchemist as execution platform for Protelis
Alchemist [PMV13] is a simulator inspired by stochastic chemistry
Features (the subset of interest for this tutorial)
Support for executing Protelis programs
Several types of environments:
Euclidean spaces, possibly with obstacles (for indoor environments)
Real world maps (using Openstreetmap data), support for GPS traces
YAML [BKEdN01] as speciﬁcation language
Multiplatform (hosted on Java 8+)
Already in use for
Crowd evacuation [MVR+
11]
Crowd steering [BPV15]
Pervasive computing [MPV12a, SYD+
13, PBV16, BVPD16, VBPB16, PDV17]
Biochemistry [MPV12b]

images/logo
Simulation examples

images/logo
Workbench setup I
Let’s have fun together!
Experimenting is the best way of learning a new paradigm and/or language
Prerequisites
Internet connection
Though I might have prepared a stand-alone version on a USB drive
A working installation of Java 8+
A terminal emulator
Unix-like terminals are ok
Should work also with the Windows cmd.exe
Don’t ignore case sensitivity
git (preferred), or a program to decompress a zip ﬁle

images/logo
Workbench setup II
Download
In the following, WORKSPACE represents the folder you want to work
into
Option 1 – git is installed on the system. Issue:
git clone
https://github.com/AlchemistSimulator/SASO17Tutorial.git
WORKSPACE
cd WORKSPACE
Option 2 – git???
Download from: http://bit.ly/saso17tutorial-zip
Unzip in WORKSPACE
Open a terminal in WORKSPACE

images/logo
Workbench setup III
First run
Issue the following command:
On Unix terminals:
./gradlew -Pfile=nothing
On Windows Prompt:
gradlew.bat -Pfile=nothing
Expectation:
Lots of downloads
More downloads
A graphical interface with a white background

images/logo
Setup explained I
The Gradle mini-box
Gradle is a (very trendy) build system based on Java and Groovy. We use
it to:
Download the software automagically
Execute the simulator
It’s easier, quicker and more portable than using sh or java directly
It reads what to do from the build.gradle ﬁle

images/logo
Setup explained II
build.gradle
// We will need Java
apply plugin: 'java'
// All the stuff can be found on Maven Central
repositories { mavenCentral() }
// We need the simulator at version 6.0.2 (change the number to test with newer versions!)
dependencies { compile "it.unibo.alchemist:alchemist:7.0.0" }
// OK that's a bit more complicated, but in the end just configures a Java process and launches it
task "run$file"(type: JavaExec){ // Create a new task with a dynamically discovered name (Groovy coolness)
classpath = sourceSets.main.runtimeClasspath // Put the simulator and its dependencies in the classpath
classpath 'protelis' // Put all the protelis files in the classpath (so the interpreter can find them)
classpath 'maps' // Cool stuff for the last example :)
main = 'it.unibo.alchemist.Alchemist' // Launch this java class
args(
"-y", "sim/${file}.yml",
"-g", "effects/${file}.aes",
"-e", "exported-data"
) // With parameters "-y sim/<SOMETHING>.yml -g effects/<SOMETHING>.aes -e exported-data"
}
/*
* If nothing different is specified, search for a property named "file" in the project, then run a task
* with such name.
*/
defaultTasks "run$file"

images/logo
A simulated environment
Let’s displace 500 nodes in a circle in a bidimensional euclidean space.
00-devices-in-a-circle.yml
incarnation: protelis
displacements: # In this section we list our device displacements
- in:
type: Circle
parameters:
- 500 # Number of devices
- 0 # X center of the circle
- 0 # Y center of the circle
- 50 # radius
Run with ./gradlew -Pfile=00-devices-in-a-circle
Expected: 500 devices (black dots) randomly displaced inside a circle

images/logo
Pan, zoom, node information
Let’s play a little bit with the graphical interface:
Pan by left click + drag
Zoom with the mouse wheel
Rotate by right click + drag
The node closest to the mouse is marked with a yellow and red dot
Double click to open a frame with the node information
Position
Content
Programs to execute
Press M on your keyboard to disable the marker
Useful e.g. if you want to grab a screenshot

images/logo
Manually moving nodes around
Nodes can be moved around by hand:
The S key enters and exits the “selection mode”
Enter in selection mode, select a group of nodes by dragging a
rectangle around them
Press O
Drag the nodes to another position
Be wary that:
Moving around is very useful for demoing and prototyping
Moving stuﬀ manually around will make your experiment non
reproducible
We’ll see other means of moving nodes and retain reproducibility

images/logo
Drawing on screen in Alchemist I
Alchemist supports the definition of a stack of graphic effects in order to
use colors and shapes to understand what’s going on.
Saved as a JSON file, that can be manually edited before launching
the simulation
The effects can be edited and saved from within the GUI
The stack is applied one effect at a time
Upper layers cover lower layers
With some training, it’s possible to achieve nice results using
transparencies
The effect system, and the whole GUI, is on a process of heavy
reworking
We are switching to the modern JavaFX
We are making the effects much easier to create and manipulate from
the UI
The plan is to release it by the end of 2017, stay tuned

images/logo
Drawing on screen in Alchemist II
00-devices-in-a-circle.aes
[
{
"type": "class it.unibo.alchemist.boundary.gui.effects.DrawShape",
"curIncarnation": "protelis",
"mode": "FillEllipse",
"red": {
"max": 255,
"min": 0,
"val": 0
},
"blue": {
"max": 255,
"min": 0,
"val": 0
},
"green": {
"max": 255,
"min": 0,
"val": 0
},
"alpha": {
"max": 255,
"min": 0,
"val": 255
},
"scaleFactor": {
"max": 100,
"min": 0,
"val": 50
},Pianini, Viroli, Beal (UniBo, BBN) Programming with Protelis 2016-09-22 – SASO Tutorial 26 / 93

images/logo
Network model
If we want the node to connect to each other, we must tell the simulator
how to do so:
01-devices-connected.yml
# Connect nodes that are sufficiently close
network-model:
type: EuclideanDistance
parameters: [10]
displacements:
# Expected: a YAML list of device displacements
- in:
type: Circle
parameters: [500, 0, 0, 50]
Launch with ./gradlew -Pfile=00-devices-connected
Expected: same as the previous experiment
Press L on the keyboard to visualize the network links!
Try to move them (S then O then drag and drop)

images/logo
On the type ... parameters syntax
Alchemist can load any implementation of its interfaces
For each type of object written in the simulation file, Alchemist tries to find and build a
Java object implementing a concrete version. The resolution is operated as follows:
1 if type and (optionally) parameters are found as keys of the object, a class with
the corresponding name is searched and a constructor is invoked with the provided
parameters
All parameters that can be deduced by the context are automatically bound
(objects of type Environment, RandomGenerator, Incarnation, Node,
Reaction, LinkingRule) and must not be provided in the parameters list
2 If the provided parameter is a string, it is forwarded to the current Incarnation
for interpretation
3 In any other case, a default system value is looked for
If the lookup or the build fail, the simulation won’t get built.
To find out the available types and their constructors, refer to the Javadoc at:
http://alchemist-doc.surge.sh/

images/logo
My ﬁrst Protelis program I
Now press P to begin the simulation
Uh?! Nothing happens... The time just jumps to Infinity.
In fact, there are no events in this simulation :)
Let’s program the nodes with an extremely simple Protelis program: 0

images/logo
My ﬁrst Protelis program II
02-0.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
- # This list of lists seems useless, it actually makes sense when you use YAML anchors
# Round frequency (advanced distributions can be used with the time/parameters syntax)
- time-distribution: 1
# This is the program we want to execute locally
program: 0
# This tells the simulator to share the results of the last computation. No time distribution has
``as soon as possible'' semantics (Markovian event with inifinite rate). The separation is
useful in case the user wants to simulate a delay between the computation and the delivery
of messages, e.g. to simulate network delays, or a battery saving policy on the wireless
modules.
→
→
→
→
- program: send
Run with: ./gradlew -Pfile=02-0
Then press P to begin the simulation

images/logo
My first Protelis program III
Expected:
The time flows!
Nodes turn red
I’ve configured an effect that changes color if the node has computed,
it’s no black magic
If you open a window with the node contents:
In “Contents”, 0 is associated to default module:default program
It’s a Protelis-generated name for anonymous modules
In “Program”, two entry appear: one is the round computation, the
other one is the event sending information to neighbors
You can use P to pause and restart the simulation

images/logo
Speeding up the simulation in Alchemist
Improve the performance
If you attached a proﬁler to Alchemist, you’ll see that most of the
time is spent in rendering nodes
By default, the simulator renders the entire scene (within the
viewport) after every event
This can get very expensive very quickly
You can tune the refresh rate of the UI by pressing:
Right arrow: speed up (less calls to the graphics)
Left arrow: speed down (more calls to the graphics)
If you speed up too much, the UI will become sluggish (e.g. if you
update it only twice per second)
R can be used to try to make the time ﬂow linearly
Useful in case of sparse, unevenly distributed events
Left and right arrows keep working
Lots of improvements are down the line, but not there yet :(

images/logo
Loading a Protelis file I
Writing Protelis specification within a YAML file is not what we want
No portability
Easy to make mistakes
Horrible kludges to make YAML happy with pieces of Protelis inside
Maybe you want to run your code elsewhere, not just inside a simulator
Let’s load a Protelis program from its own file!
tutorial/zero.pt
// Make sure that the module is in the classpath!
module tutorial:zero
0 // Program script

images/logo
Loading a Protelis ﬁle II
03-load-module.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: tutorial:zero # Just use the name of the module! The loading
happens transparently from the classpath (which, of course, means
that your Protelis source folder should be included in your Java
classpath)
→
→
→
- program: send
Run with: ./gradlew -Pfile=03-load-module

images/logo
Core language constructs
Outline
1 Introduction
3 Libraries
Making nodes move

images/logo
Core language constructs The boring stuﬀ: functions, operators, and the like
Factorial in Protelis I
tutorial/factorial.pt
module tutorial:factorial
/*
* The language is functional, every expression has a return value. In case of multiple statements in a
block, the value of the last expression is returned.→
*
* Comments are C-like, both single and multiline supported.
*
* The following defines a new function. If the optional "public" keyword is present, the function will be
accessible from outside the module→
*/
public def factorial(n) { // Dynamic typing
if (n <= 1) {
1 // No return keyword, no ";" at the end of the last line
} else { // else is mandatory
n * factorial(n - 1) // infix operators, recursion
}
}
// There is no main function, just write the program at the end (Python-like)
let num = 5; // mandatory ";" for multiline instructions
factorial(5) // Function call

images/logo
Factorial in Protelis II
04-factorial.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: tutorial:factorial
- program: send
Run with: ./gradlew -Pfile=04-factorial

images/logo
Factorial in Protelis III
Expected result
All network nodes independently compute the result
Those which computed get a purple circle around
The value of the computation gets written aside them
Again, no magic. I configured an effet to do so.
Mastering the effect system of Alchemist is not trivial...
and maybe not worth doing it, as it will get updated soon
...but may provide nice insights on the system
and nice screenshots on your papers

images/logo
Higher order, lambdas, tuples, Java interoperability I
tutorial/randomfactorial.pt
module tutorial:randomfactorial
import java.lang.Math.random // Import java static methods
def dice() {
let rand = random(); // call methods as if they were local functions
(6 * rand).intValue() // Java-style method invocation on target objects
}
def factorial(n) {
if (n <= 1) { 1 } else { n * factorial(n - 1) }
}
def callWithDice(fun) { // Higher-order function
let num = dice();
fun.apply(num)
}
[ // Use square brackets to build a tuple
callWithDice(factorial), // higher-order call
callWithDice((n) -> {n ^ 2}) // as above, but with an anonymous function (square)
] * [ 1, 2 ] // Tuple can be used with operators (they must have the same length though)
/* Elements of a tuple can be accessed using get(), several methods allow to create new tuples using
existing ones. Refer to this javadoc: http://protelis-doc.surge.sh/ */→

images/logo
Higher order, lambdas, tuples, Java interoperability II
05-higherorder.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: tutorial:randomfactorial
- program: send
Run with: ./gradlew -Pfile=05-higherorder

images/logo
Accessing the device, sensing, actuating I
tutorial/deviceaccess.pt
module tutorial:deviceaccess
/*
* Two important entities:
* - self: provides access to the device implementation. Different devices have different capabilities, so
some Protelis code may not be able to run on some platforms (e.g. code requiring access to the
current coordinates may not work in a device implementation that does not expose such information).
See ExecutionContext and its sub-interfaces at
http://protelis-doc.surge.sh/org/protelis/vm/ExecutionContext.html
→
→
→
→
* - env: provides access by name to sensors/actuators. They are treated uniformly as variables, similarly
as accessing a map→
*/
let injected = if (env.has("value")) { env.get("value") } else { "No value" }; // Read a value, if present
env.put("a generated value",
(100 * self.nextRandomDouble()).intValue()); // This is the best way to ask the platform for a random
self.getDeviceUID().getId() // The unique identifier of this device

images/logo
Accessing the device, sensing, actuating II
06-deviceaccess.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: tutorial:deviceaccess
- program: send
contents: #in this section we can pre-configure the value of sensors
- in: # Optionally, we can limit the area of space where to inject such value
type: Rectangle
parameters: [0, 0, 100, 100]
molecule: value # name of the sensor
concentration: "somevalue" # initial value (any type)
Run with: ./gradlew -Pfile=06-deviceaccess

images/logo
Core language constructs Stateful computation with rep
Evolving fields with time I
We only computed in a stateless fashion
No, using env as a mean to retain state is not a good idea
In field calculus, evolving fields are built with the rep operator
rep
0
(x)=>x+1
t
v0
t
v1
..
rep(0){(x)=>x+1}

images/logo
Evolving ﬁelds with time II
Syntax and semantics in Protelis
rep (value <- initial) { body }
When rep is encountered for the ﬁrst time, value is set to initial
Then body is evaluated, and the result stored in value
So, in the next round, value will have the result of the evaluation of
body in the previous round

images/logo
A reset counter I
primitives/counter.pt
module primitives:counter
public def cyclicCounter(min, max) {
// Let's start from min - 1, so that at the first round we get min.
rep(count <- min - 1) {
// mux is a functional multiplexer: evaluates both branches and returns one
of them→
mux(count >= max) { // Here == would suffice
0
} else {
count + 1
}
}
}
cyclicCounter(0, 10)

images/logo
A reset counter II
07-counter.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: primitives:counter
- program: send
contents:
Run with: ./gradlew -Pfile=07-counter

images/logo
Using diﬀerent time distributions
08-async.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
- time-distribution: # Check out in the Alchemist Javadocs all the classes implementing
"TimeDistribution" if you want a complete overview of what's available→
type: ExponentialTime # Markovian-distributed events with rate = 1
parameters: [1]
program: primitives:counter
- program: send
contents:
Run with: ./gradlew -Pfile=08-async
Devices run with a random (and memoryless) time distribution
(variance equal to the mean): very chaotic behavior.

images/logo
Core language constructs Sharing information with nbr
Distribution: view of the field in the local neighborhood I
We did not yet use the communication abilities of our devices
It’s time to do it!
In field calculus, fields of fields are built with the nbr operator
It builds a field of views of a field in the neighborhood
nbr de
nbr{e}
φd=[d1→v1,..,dn→vn]

images/logo
Distribution: view of the ﬁeld in the local neighborhood II
nbr(body)
Evaluates body
Returns a ﬁeld containing the mapping between the neighbouring
aligned devices (included self) and the value of body at their location
A device is d0 aligned to another device d1 if, recently enough, d0
executed the same nbr instruction d1 is executing, sharing the result
with it.
Shares the local value of body

images/logo
Reduction and manipulation of fields of fields I
Reducing fields
The information on a field is usually reduced to a single value (a field,
not a field of fields)
Such operation is performed by the hood built-in function
hood(reduction, defval, fieldExpr)
hood PlusSelf(reduction, defval, fieldExpr)
Where reduction is a function of two parameters (T, T)->T, defval is
a value of type T, and fieldExpr a field of fields of T
fieldExpr is evaluated
if PlusSelf is not specified, the local value is discarded
reduction is applied to every value
If there is a result it is returned, defval si returned otherwise

images/logo
Reduction and manipulation of fields of fields II
Built-in hoods
Protelis has a number of hood operations ready to use (every one with its
PlusSelf variant)
minHood, minimum of a field, default Infinity
maxHood, maximum of a field, default -Infinity
anyHood, “or” on a field of booleans, default false
allHood, “and” on a field of booleans, default true
sumHood, sum of the values of the field, default 0
meanHood, mean of the values of the field, default NaN
unionHood, a tuple with all the values in the field, default []

images/logo
Reduction and manipulation of fields of fields III
Manipulation of fields of fields
Fields and locals (using the global view: fields of fields and fields) can
be used in any operation
e.g. nbr(v) + 1 will produce as result a field containing for every
neighbor the value of v + 1
e.g. nbr(v) + nbr(a) will combine the values neighbor-per-neighbor
it also works for method call: nbr(anObject).hashCode() will return a
field of hash codes

images/logo
Reduction and manipulation of fields of fields IV
Manipulation and alignment
Manipulation of fields of fields is possible because the two instructions
that can “break the alignment” (rep and if—to be seen soon)
cannot return fields of fields
Fields of fields are guaranteed to always have the same devices at
every step of the execution!

images/logo
Pace-making counter I
primitives/pacemaker.pt
module primitives:pacemaker
// Import public functions from another module
import primitives:counter
def max(a, b) { // a max working on anything that is comparable
if (a > b) { a } else { b }
}
def counter() { rep (i <- -1) { i + 1 } } // plain counter
public def pacemaker(min, max) {
let myCount = [counter(), cyclicCounter(min, max)]; // tuple with the count of my rounds and the value of
the cyclic counter→
rep (count <- myCount) { // pick and store...
maxHood PlusSelf(nbr(max(myCount, count))) // ...the maximum in my neighborhood between the maximum of
my count and the maximum i got until now→
}.get(1) // we are interested in the second argument of the tuple, the first an implementation detail
}
pacemaker(0, 10)

images/logo
Pace-making counter II
09-pacemaker.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
- time-distribution:
type: ExponentialTime
parameters: [1]
type: Event # We define the event manually...
actions: # ...because we want to specify a parameter of the Protelis interpreter.
- type: RunProtelisProgram
parameters: ["primitives:pacemaker", 5] # We discard messages after 5 simulated seconds
- program: send
Run with: ./gradlew -Pfile=09-pacemaker
The fastest device generates a “wave” that pulls up the slow ones!

images/logo
Gradient: computing distances I
primitives/gradient.pt
module primitives:gradient
/*
* Compute the distance to the closest source. The input is a field of booleans, this function computes the
distance from the closest device where the field yields true.→
*/
public def gradient(source) {
rep (d <- Infinity) { // Start from an infinite distance
mux (source) { // Functional multiplexer (evaluates both branches)
0 // If I am the source, my distance is zero
} else {
minHood(nbr(d) + self.nbrRange()) // The distance is computed as the minimum of the sum of the
distance between me and my neighbors, and my neighbors and the source, for each neighbor.→
}
}
}
// Let's compute the distance from the device "zero"
gradient(self.getDeviceUID().getId() == 0)

images/logo
Gradient: computing distances II
10-gradient.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: primitives:gradient
- program: send
Run with: ./gradlew -Pfile=10-gradient
The fastest device generates a “wave” that pulls up the slow ones!

images/logo
Gradient: computing distances I
primitives/gradient.pt
module primitives:gradient
/*
* Compute the distance to the closest source. The input is a field of booleans, this function computes the
distance from the closest device where the field yields true.→
*/
public def gradient(source) {
rep (d <- Infinity) { // Start from an infinite distance
mux (source) { // Functional multiplexer (evaluates both branches)
0 // If I am the source, my distance is zero
} else {
minHood(nbr(d) + self.nbrRange()) // The distance is computed as the minimum of the sum of the
distance between me and my neighbors, and my neighbors and the source, for each neighbor.→
}
}
}
// Let's compute the distance from the device "zero"

images/logo
Gradient: computing distances II
10-gradient.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
- program: send
Run with: ./gradlew -Pfile=10-gradient
Move the source node around and see what happens!

images/logo
Core language constructs Domain restriction with if
Domain restriction with if I
In aggregate programming, using if has diﬀerent consequences than
the languages you are used to
It’s not a “ﬂow control” instruction
It’s a domain restriction instruction
Devices get separated in two, non communicating, domains
They lose alignment
Reasoning at the device level can be very misleading!
if
eb
if
cond
then
else
if (eb) {e} else {false}
false
e

images/logo
Domain restriction with if II
if(condition) { then } else { otherwise }
Evaluates condition
Executes either then or otherwise, depending on the value of condition
Unexecuted branches lose their state
reps get reset to default
If one of the branches contains:
nbr
rep
a call to a function using the two above
then the choice between if and mux is paramount!

images/logo
Screw up the gradient I
primitives/brokengradient.pt
module primitives:brokengradient
def gradient(source) {
rep (d <- Infinity) {
if (source) { // Changed mux with if
0
} else {
// The source never enters this domain!
// Frome the single device perspective, it never executes this branch, and so it never shares its
knowledge of being a source!→
minHood(nbr(d) + self.nbrRange())
}
}
}

images/logo
Screw up the gradient II
11-brokengradient.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: primitives:brokengradient
- program: send
Run with: ./gradlew -Pfile=11-brokengradient
Our toy got bricked!

images/logo
Mixing rep, if and mux I
primitives/twocounters.pt
module primitives:twocounters
def counter() { rep(c <- 0) { c + 1 } }
[
if (counter() % 2 == 0) { counter() } else { counter() },
mux (counter() % 2 == 0) { counter() } else { counter() }
]
12-twocounters.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: primitives:twocounters
- program: send

images/logo
Mixing rep, if and mux II
Run with: ./gradlew -Pfile=12-twocounters
if continuously resets the rep status!

images/logo
Gradient in a sub-domain I
primitives/gradobs.pt
module primitives:gradobs
import primitives:gradient
if (env.has("obstacle")) {
Infinity
} else {
}

images/logo
Gradient in a sub-domain II
13-gradobs.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: primitives:gradobs
- program: send
contents:
- in:
type: Rectangle
parameters: [-15, -50, 30, 75]
molecule: obstacle
concentration: true
Run with: ./gradlew -Pfile=13-gradobs
Modify the program to use mux, and see the diﬀerence!

images/logo
Libraries
Outline
1 Introduction
3 Libraries
Making nodes move

images/logo
Libraries Frequently used patterns: towards reusable building blocks
Raise the abstraction level
Programming in terms of rep, nbr and if is very, very low level
And also, very much error-prone
What would happen if minHood PlusSelf was used in the gradient?
Try it in the simulator, modifying the 10-gradient example
Move nodes around: it’s not self-stabilizing anymore...
Several coordination patterns are common to many algorithms
Spread information
Accumulate information
Evolve state with time
Partition the system
We want to program with them directly
There are several ways of implementing such primitives
With diﬀerent properties, both functional and non-functional

images/logo
The G / C / T / S building blocks I
G: spreading information
G(source, initial, metric, accumulate)
Spreads initial from source
initial can be of any type (T)
source is a boolean
metric is a function ()->Field<Number> that computes the gradient variation (the same
role played by nbrRange in our gradient)
values are combined with accumulate, a function (T)->T
Example: rewrite the gradient with G
def distanceTo(source) {
let metric = () -> { self.nbrRange() };
G(source, 0, metric, v -> { v + metric.apply()})
}

images/logo
The G / C / T / S building blocks II
C: accumulate information
C(potential, reduce, local, null)
Accumulates local values along a potential
potential is a Comparable type (typically a number)
reduce is a (T, T) -> T
local is the local value (T)
null is the default value, in case there is nothing to accumulate
Example: Count the devices in the system, accumulating the information in source
def countDevices(source) {
C(distanceTo(source), sum, 1, 0)
}

images/logo
The G / C / T / S building blocks III
T: evolve state
T(initial, zero, decay)
A timer from initial to zero with a decay rate.
S: partition the network
S(grain, metric)
Elects a leader every grain, using the provided metric.
Returns a boolean (true if the node is a leader)
grain is a number with the partition size
metric is a ()->Field<Number> that computes the distances
A timer from initial to zero with a decay rate.

images/logo
Multiple gradients I
advanced/voronoi.pt
module advanced:voronoi
import protelis:coord:sparsechoice
import protelis:coord:spreading
G(S(25, metric), 0, metric, v -> { v + metric.apply() })
It’s one line of “true” code

images/logo
Multiple gradients II
14-voronoi.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: advanced:voronoi
- program: send
Run with: ./gradlew -Pfile=14-voronoi
Try to move the nodes around, and see how the system adapts!

images/logo
Libraries Filling the abstraction gap: Protelis-lang
Limitations of the building block
Programming with building blocks is way easier than ﬁddling with
reps and nbrs by hand
...but it is not yet where we’d like to stand!
Many “useful” behaviors can’t be captured
as they are not self-stabilizing
Yet, they can be of use in many situations
Think of gossip, for instance
It’s very hard to get a grasp of the global situation
If there is an information spread from multiple sources, only
information from the closest would make it
This is ok in many situations
But not enough in others
Many algorithms recur very often
They should stay in a library

images/logo
Use protelis-lang and focus on the application code
protelis-lang is a library1 intended to:
Provide a collection of the most widely used distributed algorithms
Contains over a hundred of them, from literature and other tools
If you develop something new, contribution is very welcome :)
Shift the focus of the designer entirely on application code
Provide means to overcome the limitations of the building blocks
Through a number of meta-algorithms that allow parallel aligned
executions
Mainly by means of the novel multiInstance algorithm
protelis-lang is part of the Protelis distribution and ready to use
1
First introduced this year at a SASO Workshop: the 2nd edition of eCAS — did you
follow the presentation? :)

images/logo
Partition the network and spread multiple gradients I
advanced/voronoi.pt
module advanced:voronoi
G(S(25, metric), 0, metric, v -> { v + metric.apply() })
It’s one line of “true” code

images/logo
Partition the network and spread multiple gradients II
advanced/summarize.pt
module advanced:summarize
import protelis:lang:utils
import protelis:coord:accumulation
import protelis:state:time
// Let's simulate a sensor measuring the temperature (in Celsius, range 20 to 30) every minute
env.put("temp", rep (v <- 25) { if(cyclicTimer(60)) { 20 + self.nextRandomDouble() * 10 } else { v }});
// Let's compute the mean temperature of the whole network and share th infomation to all devices
let sink = S(Infinity, nbrRange);
summarize(sink, sum, env.get("temp"), 0) / broadcast(sink, countDevices(distanceTo(sink)))

images/logo
Partition the network and spread multiple gradients III
15-summarize.yml
network-model:
parameters: [10]
displacements:
- in:
type: Circle
parameters: [500, 0, 0, 50]
programs:
-
program: advanced:summarize
- program: send
Run with: ./gradlew -Pfile=15-summarize

images/logo
More about the toolchain
Outline
1 Introduction
3 Libraries
Making nodes move

images/logo
More about the toolchain Making nodes move
Brownian movement I
16-brownian.yml
network-model: { type: EuclideanDistance, parameters: [10] } # Compact syntax
program-pools:
- compute-gradient: &gradient #This is a YAML anchor: assigns a name to this YAML object, and can be
reused later in the file. It's not an Alchemist thing, it's pure YAML specification→
- program: send
- move: &move
- time-distribution: { type: ExponentialTime, parameters: [1] }
type: Event
actions:
- { type: BrownianMove, parameters: [1] }
displacements:
- in: { type: Circle, parameters: [500, 0, 0, 50] }
programs:
# Now maybe the strange double list in "programs" makes sense :)
- *gradient # Reference to the anchor previously defined.
- *move
Run with: ./gradlew -Pfile=16-brownian

images/logo
Controlling the movement direction I
advanced/converge.pt
module advanced:converge
import protelis:state:time
let leader = S(25, nbrRange);
let destination = broadcast(leader, self.getCoordinates());
destination = if (isSignalStable(destination, 10)) { destination } else { self.getCoordinates() };
env.put("destination", destination);
self.getCoordinates() == destination

images/logo
Controlling the movement direction II
17-converge.yml
network-model: { type: EuclideanDistance, parameters: [10] }
program-pools:
- compute-gradient: &gradient
program: advanced:converge
- program: send
- move: &move
type: Event
actions:
- { type: MoveToTarget, parameters: [destination, 1] }
displacements:
- in: { type: Circle, parameters: [500, 0, 0, 50] }
programs:
- *gradient
- *move
Run with: ./gradlew -Pfile=17-converge

images/logo
More about the toolchain Conﬁguring batches and exporting data
Exporting data I
18-export.yml
network-model: { type: EuclideanDistance, parameters: [10] }
program-pools:
- { time-distribution: 1, program: "advanced:converge" }
- program: send
- move: &move
type: Event
actions: [ { type: MoveToTarget, parameters: [destination, 1] } ]
displacements: [ { in: { type: Circle, parameters: [500, 0, 0, 50] }, programs: [*gradient, *move] } ]
export:
- time # Exports the current time
- number-of-nodes # Exports the number of nodes in the system
- molecule: advanced:converge # The name of the sensor / actuator whose value will be exported
# Optionally, a "property: program" section can be specified, where "program" is a valid chunk of
protelis code. The sensor value will be processed using such code before being exported.→
value-filter: onlyfinite # You may want to keep poisonous values (Infinities and NaN) from being
exported or passed to the aggregator. Available filters are "nofilter" (default), "onlyfinite"
(discards both NaN and Infinities), "filternan", "filterinfinity".
→
→
aggregators: [sum, mean, kurtosis] # The aggregator takes all the data from nodes and reduces it to a
single value. If omitted, one value per node will be exported (in this case, 500 columns...). You
can load any UnivariateStatistics from the Apache Commons Math library, just by their name.
→
→

images/logo
Exporting data II
Run with: ./gradlew -Pfile=18-export
Look at the ﬁle exported-data.txt

images/logo
Using variables and controlling randomness I
19-variables.yml
variables:
seed: &seed # You can give the anchor any name, assigning the name of the variable is convenient, though
{min: 0, max: 9, step: 1, default: 0} # This variable ranges in [0, 9], steps of 1, defaulting to 0
connection-radius: &connection-radius
formula: 10 # An instance of a Javascript interpreter is created to evaluate the expression.
round-frequency: &round-frequency
formula: Math.max($seed, 1) # Other variables can be referred by prefixing them with £ (e.g. £seed would
be substituted by the current value of seed).→
language: scala # The system defaults to Javascript, but a Scala interpreter can be instanced instead
program-name: &program-name
formula: "'advanced:converge'" # Use whatever type you want - just make sure you use them properly
seeds:
scenario: *seed # This controls the initial displacement of the nodes
simulation: *seed # This controls the generated event times
network-model: { type: EuclideanDistance, parameters: [*connection-radius] }
program-pools:
- { time-distribution: 1, program: *program-name }
- program: send
- move: &move
- time-distribution: { type: ExponentialTime, parameters: [*round-frequency] }
type: Event
actions: [ { type: MoveToTarget, parameters: [destination, 1] } ]
displacements: [ { in: { type: Circle, parameters: [500, 0, 0, 50] }, programs: [*gradient, *move] } ]

images/logo
Using variables and controlling randomness II
Run with: ./gradlew -Pfile=19-variables
Run it again changing the default for the seed variable!
Variables can be instanced with the type/parameters syntax
Several implementations available (geometrically scaling, arbitrary
values...)
See http://alchemist-doc.surge.sh/it/unibo/alchemist/
loader/variables/package-frame.html

images/logo
Using variables and controlling randomness III
Batches
Independent variables (those with no formula entry) deﬁne the batch
content
if Alchemist is executed with the --batch, it will run in headless
mode, spawning a simulation for each possible combination of values
of the variables listed after the -var option
e.g.: java -jar alchemist.jar --batch --var seed range
If range may yield 1, 10 or 100 and seed may yield 0 or 1
Six simulations would get generated
Don’t abuse: with many variables and many variables thing get big
very quickly
Simulations are executed in parallel
By default one simulation per core + 1
This behavior is conﬁgurable with a command line option
A grid executor is in our TODO list

images/logo
More about the toolchain Using real world maps
Loading maps I
20-maps.yml
environment:
type: OSMEnvironment
parameters: ["/vcm.pbf"] # Loads from classpath, an absolute path to the file would work as well
pools:
- pool: &move
- time-distribution: 0.1
type: Event
actions:
- type: ReproduceGPSTrace # Other strategies are available, such as interpolate the track using
pedestrian roads→
parameters: ["/vcmuser.gpx", false, "AlignToTime", 1365922800, false, false]
displacements:
- in:
type: FromGPSTrace # displace nodes as they are in the GPS track
parameters: [1497, "/vcmuser.gpx", false, "AlignToTime", 1365922800, false, false]
programs:
- *move
Run with: ./gradlew -Pfile=20-maps

images/logo
More about the toolchain Getting from simulations to devices
Writing a backend I
Starting point: provide an ExecutionContext
Your back-end entry point is the implementation of the
ExecutionContext interface
Many complicated functions
Fortunately, you can just extend from AbstractExecutionContext
Just a single instance() method left virtual, basically a call to the
constructor.
You can see how’s implemented in the simulator for a clue

images/logo
Writing a backend II
Expose device capabilities
The ExecutionContext also holds device capabilities (it’s like the
System / Runtime singletons of Java).
A number of interfaces for standard capabilities are provided
TimeAwareDevice — Implement this if your device is able to estimate
network message delays and lags
SpatiallyEmbeddedDevice — Implement this if your device is able to
estimate distances to its neighbors (with whatever metric)
LocalizedDevice — If your device has access to its position (e.g., if a
GPS system is onboard)
You are free to add any method, and access it via calls to self
The thing usually implemented is something along the line of
MyExecutionContext extends AbstractExecutionContext implements
TimeAwareDevice, LocalizedDevice

images/logo
Writing a backend III
Implement communication
Communication is dealt with by the NetworkManager
void shareState(Map<CodePath,Object> toSend)
You are in charge of ﬁnding a way to send your neighbors the toSend
payload...
Map<DeviceUID,Map<CodePath,Object>> getNeighborState()
...and of presenting the payloads sent to others to the current device.

images/logo
Writing a backend IV
Run
Now that pieces are in place, just create a ProtelisVM and run whatever
you like, e.g.:
final NetworkManager netmgr = new MyNetworkManager(...);
final ExecutionContext ctx = new MyExecutionContext(netmgr, ...);
final ProtelisProgram prog = ProtelisLoader.parse(myProgram)
final ProtelisVM vm = new ProtelisVM(prog, ctx);
while (alive) {
vm.runCycle();
Thread.sleep(someTime);
}

images/logo
References
References I
Oren Ben-Kiki, Clark Evans, and Ingy d¨ot Net.
Yaml ain’t markup language (yaml™), version 1.2, 3rd edition, 2009-10-01.
http://yaml.org/spec/1.2/spec.pdf.
Jacob Beal, Danilo Pianini, and Mirko Viroli.
Aggregate programming for the internet of things.
IEEE Computer, 48(9):22–30, 2015.
Jacob Beal, Mirko Viroli, Danilo Pianini, and Ferruccio Damiani.
Self-adaptation to device distribution changes.
In 10th IEEE International Conference on Self-Adaptive and Self-Organizing Systems,
SASO 2016, Augsburg, Germany, September 12-16, 2016, pages 60–69, 2016.
Sara Montagna, Danilo Pianini, and Mirko Viroli.
Gradient-based self-organisation patterns of anticipative adaptation.
In Sixth IEEE International Conference on Self-Adaptive and Self-Organizing Systems,
SASO 2012, Lyon, France, September 10-14, 2012, pages 169–174, 2012.

images/logo
References
References II
Sara Montagna, Danilo Pianini, and Mirko Viroli.
A model for drosophila melanogaster development from a single cell to stripe pattern
formation.
In Proceedings of the ACM Symposium on Applied Computing, SAC 2012, Riva, Trento,
Italy, March 26-30, 2012, pages 1406–1412, 2012.
Sara Montagna, Mirko Viroli, Matteo Risoldi, Danilo Pianini, and Giovanna Di Marzo
Serugendo.
Self-organising pervasive ecosystems: A crowd evacuation example.
In Software Engineering for Resilient Systems - Third International Workshop, SERENE
2011, Geneva, Switzerland, September 29-30, 2011. Proceedings, pages 115–129, 2011.
Danilo Pianini, Jacob Beal, and Mirko Viroli.
Improving gossip dynamics through overlapping replicates.
In Coordination Models and Languages - 18th IFIP WG 6.1 International Conference,
COORDINATION 2016, Held as Part of the 11th International Federated Conference on
Distributed Computing Techniques, DisCoTec 2016, Heraklion, Crete, Greece, June 6-9,
2016, Proceedings, pages 192–207, 2016.

images/logo
References
References III
Danilo Pianini, Simon Dobson, and Mirko Viroli.
Self-stabilising target counting in wireless sensor networks using euler integration.
11th IEEE International Conference on Self-Adaptive and Self-Organizing Systems, SASO
2017, Tucson, AZ, USA, September 18-22, 2017, 2017.
To appear.
Danilo Pianini, Sara Montagna, and Mirko Viroli.
Chemical-oriented simulation of computational systems with ALCHEMIST.
J. Simulation, 7(3):202–215, 2013.
Danilo Pianini, Mirko Viroli, and Jacob Beal.
Protelis: practical aggregate programming.
In Proceedings of the 30th Annual ACM Symposium on Applied Computing, Salamanca,
Spain, April 13-17, 2015, pages 1846–1853, 2015.
Graeme Stevenson, Juan Ye, Simon Dobson, Danilo Pianini, Sara Montagna, and Mirko
Viroli.
Combining self-organisation, context-awareness and semantic reasoning: the case of
resource discovery in opportunistic networks.
In Proceedings of the 28th Annual ACM Symposium on Applied Computing, SAC ’13,
Coimbra, Portugal, March 18-22, 2013, pages 1369–1376, 2013.

images/logo
References
References IV
Mirko Viroli, Antonio Bucchiarone, Danilo Pianini, and Jacob Beal.
Combining self-organisation and autonomic computing in cass with aggregate-mape.
pages 186–191, 2016.

Practical Aggregate Programming with Protelis @ SASO2017

More Related Content

Similar to Practical Aggregate Programming with Protelis @ SASO2017 (20)

More from Danilo Pianini (20)

Recently uploaded (20)

Practical Aggregate Programming with Protelis @ SASO2017