COSINE: Collaborator Selector for Cooperative Multi Device Sensing and Computing

Multi-device sensing and computing presentation for IEEE PerCom

1. COSINE: Collaborator Selector
for Cooperative Multi-Device
Sensing and Computing
Huber Flores, AgustinZuniga, FarbodFaghihi, XinLi, Samuli
Hemminki, Sasu Tarkoma, Pan Hui and Petteri Nurmi
1
March 26, 2020, Austin, Texas, USA

2. Importance
•Many devices in proximal communication ranges
2
Source: https://www.railwaypro.com/wp/mernda-rail-extension-opens/
Source: https://www.railwaypro.com/wp/mernda-rail-extension-opens/
Train station Street Home
Source: https://www.pinterest.com/pin/484840716132563124/

3. Finding collaborators is non-trivial
•Selection of collaborators is sensitiveto the type of task
Computing -> small variance in collaborations, otherwise they are hard to schedule and can easily fail
Sensing ->require long times to become effective, and tasks should be easily resumed.
•Randomlyselection of collaborators results in obtaining unpredictable
collaboration times
•Selecting most familiarcollaborators does not provide adaptation to human
mobility characteristics
1.How to maximize duration of collaborations?
2.How to make variance in time small?
3

4. COSINE: Contributions
•New method:We present COSINE for selecting optimal collaborators with
longer and more consistent duration periods in the wild.
•New insights: We show the drawbacks of existing methods to select
collaborators and demonstrate its sensitive to characteristics of human
mobility.
•Improved benefits: We demonstrate significant improvements in energy
and performance when compared to state-of-the-art solutions.
•New applications: Our COSINE enables the development of a new type of
applications powered by collaborations, e.g., Edge intelligence, Micro data
centres, and federated learning.
4

5. 5
Location estimation
task (GPS)
IdleIdle
Reduce
redundancy
Individual
processing power
Combined
processing power
New
applications
Benefits Opportunities
Micro data-centres
(e.g., edge intelligence,
federated learning)

6. State-of-the-art solutions
6
Our work
COSINE
Collaborative frameworks
Transient infrastructures
CeeSr(Computer, 2016)
CoMon(MobiSys, 2012)
CoSense(SenSys2012)
HAGGLE (UbiComp,
2007)
HyMobi(PMC 2016)
FemtoClouds(CLOUD,
2015)
RIO (MobiSys, 2014)
UbiSpark
(PerCom@PerCrowd ,
2018)
Blend (IPSN, 2017)
SCENTS
(PerCom@PerCrowd , 2019)
Stacon
(PerCom@SmartEdge ,
2019)
2007 and below
Selecting collaborators
Finding collaborators
Mobility patterns

7. •The key insight is to quantify the regularity of encounters
between devices, such that it is possible to identify devices
based on different levels of regularity.
•Regularity = (Markov trajectory) entropy values
COSINE: Overview
7
Device to device
encounter
Different regularity
values from encounters.
Each regular figure is
associated to a duration
[in min]
Quantify regularity of
encounters
=4
=2
Ranked
candidates based
on regularity
10 20 15
1
2

8. COSINE: Quantization of measurements
Phase 1
8
•Aggregate samplesinto a
signal
Data-intensive analysis
•Quantize the signal
Reduce details while
preserving relative patterns
Prepare for regularity
extraction

9. COSINE: Extraction of regularity
Phase 2
9
•Build a Markov trajectory
entropy matrix
Quantized signal is taken as
input
•Estimate the predictability of
consistent encounters
•The higher the entropy, the more
consistent (longer duration) and
vice versa

10. COSINE: Selection of collaborators
Phase 3
10
•Derive entropy ranges with
upper and lower bounds that
depict grouping of entropy
values
•Entropy ranges are ranked
based on cardinality
Candidates are selected
according to the frequency of
their entropy range

11. COSINE: Evaluation and Results
11

12. Baselines
•Familiarity
•Permanency
•Magnitude
12

13. COSINE: Evaluation
Result:Enough regularity to model different types of encounters
13

14. COSINE: Evaluation
Result: Regularity can be used to characterize different types of
encounters in a more consistent manner
14

15. COSINE: Performance
Result:Selection of collaborators is optimal and more consistent
15

16. COSINE: Different contexts
Result: Our approach adapts to different characteristics of human
mobility
16
Park Train station

17. COSINE: Energy saving
17Samsung S5 Google Nexus
Idle
Idle

18. Summary and conclusions
•New method:We present COSINE for selecting optimal collaborators with
longer and more consistent duration periods in the wild.
•New insights: We show the drawbacks of existing methods to select
collaborators and demonstrate its sensitive to characteristics of human
mobility.
•Improved benefits: We demonstrate significant improvements in energy
and performance when compared to state-of-the-art solutions.
•New applications: Our COSINE enables the development of a new type of
applications powered by collaborations, e.g., Edge intelligence, Micro data
centres, and federated learning.
18

19. Questions?
Thank you! (Do not hesitate to reach us via e-mail)
Huber Flores ([email protected])
AgustinZuniga([email protected])
FarbodFaghihi([email protected])
XinLi ([email protected])
Samuli Hemminki ([email protected])
Sasu Tarkoma([email protected])
Pan Hui ([email protected])
Petteri Nurmi ([email protected]) 19
IdleIdle
Reduce redundancy