Vector Programming Thoughts

Are all supported applications only of the periodic form. This abstraction seems to only allow code to run in periodic stages. This may be sufficient, but we need to show that.

MedEvac (and others): Would it make more sense to alert the medic when the soldier needs them rather than querying periodically? What overhead are we paying?

How costly are the operations. We should be able to place an expense on each. For example: Reading a distributed array: N messages
Find ?

How do we program and deal with special routing and network configurations. Such as the oil pipeline example from class?

Research Meeting

Thoughts:

• MPI (OpenMP)
• Parallel Compiler/Translator
• Concurrent Programming (Theads)
• Shared Variable State
• 2-Phase Program

Shared vs. Distributed Memory Model

Node-Dependent

• Karios (Pleiades v2) (Serialized parallel loops)
• Regiment
• Split-C

Node-Independent

• TAG
• Cougar
• DFuse

I do not think the program needs to run completely within the network.  We may be able to use a central controller.

From Meeting:

Pleiades

• Messaging Overhead of Control Transfer
• N messages to start and N messages to end (Parallel Threads)

Solution:

• Time binding of message response
• Threading (Matlab/Octave)

Decomposing code is key and with our application specific domains, this may be possible.

Things to look into:

• MagnetOS
• Tamim's existing applications
• Matlab code in repository
• Alternatives to Decomposing Code

Things to do:

• Matlab Abstraction (Vector Programming)
• Compiling/Translating the matlab code

1-1 Meeting Notes

1. Placelab Data Examination
1. Nothing easily
2. Prior work to IPSN
3. SVD system as model-less classifier

• Programmer tuned algorithm
• Expressiveness
• Supervised learning - Tune for all deployments
• Unsupervised learning - Tune once
• What changes for the deployment between houses?
• Time window sizes
• Time window overlap
• Number of features
• Number of clusters
• Number of noise clusters
• Ignore specific sensors

Main problem:
Input:

• 3 kinds of events with time scales
• Looking for 80+% accuracy by changes in Feature Vector

Distinguishing Component:

• Generic feature library
• Spacial features
• Temporal features

XXX: Stream Clustering for Wireless Sensor Data
(fill in XXX with name)

Hypothesis: What Features

• Pairwise
• Order by distance
• Limit the number of combinations
• Groupwise

Anthony Wood - WSN Presentation

Human in the Loop: Distributed Datat Streams for Immersive Cyber-Physical Systems

WSN limitations

1. institutional/research use
2. domain experts
3. disconnected from each other
4. privileged users
5. a priori knowledge of capabilities
6. database abstraction

Future: Democratization of sensing

Immersive CPS

1. Interactive environment
2. Overlay background data collection
3. User directed sensing

Immersive CPS Properties

1. Distributed (Sensing and Access)
2. Open (No central controller
3. Interactive (Users drive operation)
4. Discoverable (Self-advertising, self-describing)
5. Stochastic (Unpredictable demand)
6. Heterogenous (Apps, UIs, sensor, actuators, ...)
7. Localized (People are geo-physically constrained)

Research Questions:

1. Trust of sensors
   

The physical nature of the system distinguishes CPS from WSNs. Guarantees made in the system lead to CPS from WSNs.