An overview of monitoring techniques used on the edge to lower big data and energy efficiency barriers for IoT. To achieve this we introduce the AdaM and ADMin frameworks. This presentation is from a talk given at the University of Cyprus (March 2017). If used, please cite one of the following:
- "Adam: An adaptive monitoring framework for sampling and filtering on IoT devices", D. Trihinas et al., IEEE BigData 2015, 10.1109/BigData.2015.7363816
- "ADMin: Adaptive Monitoring Dissemination for the Internet of Things", D. Trihinas et al., IEEE INFOCOM 2017, to appear
Low-Cost Approximate and Adaptive Monitoring Techniques for the Internet of Things
1. 3/13/2017 1Demetris Trihinas 1
Low-Cost Approximate & Adaptive Monitoring
Techniques for the Internet of Things
Demetris Trihinas
trihinas@cs.ucy.ac.cy
PhD Candidate
Department of Computer Science
University of Cyprus
2. 3/13/2017 2Demetris Trihinas 2
Cloud Computing
Big Data Systems and Analytics
Internet Computing and IoT
Online Social Networks
Activity Areas
Current EU-Funded
Projects
http://linc.ucy.ac.cy/
3. 3/13/2017 3Demetris Trihinas 3
The Internet of Things
The physical world is now becoming an information system
Exchanging continuous data streams with other network-
enabled devices, systems and humansā¦
Physical (battery-powered) and network-enabled devices
with smart processing capabilitiesā¦
4. 3/13/2017 4Demetris Trihinas 4
Why is IoT Just Now Becoming a Reality?
Nano Bio-Sensors and
Printed Electronics
Communication
Protocols
Data-Mining
Learning Algorithms
- Distributed closed neighbors
- Outlier/event detection
- Pattern matching
- Stream processing
- Learning algorithms
- ā¦.
āThe Emerging Internet of Things Market Perspective: A Surveyā, Perera et al., IEEE Trans. Computing, 2015
āInternet of Things (IoT): A Vision, Architectural Elements, and Future Directionsā, J. Gubbi and R. Buyya, FGCS, 2013
The Building Blocks of the Internet of
Things
5. 3/13/2017 5Demetris Trihinas 5
IoT in Agriculture
Precision Farming
Detect crop condition and spread
nutrients to parts of field with need
Cattle as a Service
Detect optimum breading
period, monitor health and
dairy protection
āConnected Cattle: Wearables are Changing the Dairy Industryā, Kathy Pretz, IEEE Magazine: the Institute, May 2016.
6. 3/13/2017 6Demetris Trihinas 6
IoT in Medicine and Health Tracking
Precision Medicine
Pill-shaped cameras traversing
human body
āFitbit data led doctors to shock a patientās heartā, Steven Dent, Engadget, Apr 2016.
Wearables
Biosignal and activity tracking
7. 3/13/2017 7Demetris Trihinas 7
IoT in Intelligent Transportation Systems
Smart Traffic Flow
In Toronto travel time
was reduced by 26%
Smart Parking
In California carbon emissions
were reduced by 730 tons per
15 block radius
āCan smart traffic lights ease Torontoās road congestion?ā, Joseph Hall, Toronto Star, Jun 2014.
8. 3/13/2017 8Demetris Trihinas 8
The āBig Dataā in IoT
[IDC, Jun 2014]
[Cisco, Apr 2011]
IoT Monitoring Data
2% of digital universe in
2012
Projections for >12% in
2020
[Gartner, Mar 2015]
IP Traffic 1.6ZB in 2018, 57% from IoT Devices
[Cisco, Apr 2014]
10. 3/13/2017 10Demetris Trihinas 10
IoT and Cloud Computing
Driving Cloud Adoption for IoT Developers
- Simplified resource management
- low capital expenses to get started
[IoT Developer Survey, Eclipse Foundation, 2016]
67% of production-ready
IoT services reside in the
cloud
[State of IoT, Embarcadero, 2016]
12. 3/13/2017 12Demetris Trihinas 12
Monitoring Topologies
Scalability and Single points of failure
Monitoring servers closer to the root are overloaded
with relay costs preventing scalability
āJCatascopia: Monitoring Elastically Adaptive Applications in the Cloudā, D. Trihinas, G. Pallis and M. D. Dikaiakos, IEEE/ACM CCGrid 2014
13. 3/13/2017 13Demetris Trihinas 13
Monitoring as a Service (MaaS)
ā¢ Eases management of the monitoring infrastructure
ā¢ Service providers enable multi-tenant and scalable
monitoring over the internet
ā¢ Pay-as-you-use business model (although usually hourly)
Monitoring libraries
for metric collection from
popular languages and
frameworks
REST API to
insert/extract
monitoring data
Shared data warehouse with
services to query, process and
view analytic insights
custom app-specific metrics
managed by cloud provider
14. 3/13/2017 14Demetris Trihinas 14
Augmenting IoT with the Cloud
ā
app
The Harsh Reality
Network latencies, bandwidth limitations,
pricing (in/out cloud traffic is billed),
constant location awareness
Datacenter
Interesting Articles
āThe cloud is not enough: Saving IoT from the cloudā, B. Zhang et al., Usenix HotCloud 2015
āTaking the internet to the next physical levelā, V. Cerf and M. Senges, IEEE Computer, 2016
The IoT Developer Perspective
15. 3/13/2017 15Demetris Trihinas 15
The Cost of Monitoring
Based on AWS pricing scheme (May 2015)
for a 3-tier (video) streaming application
$0.12/GB
network traffic
āMonitoring Elastically Adaptive Multi-Cloud Servicesā, D. Trihinas, G. Pallis and M.D. Dikaiakos, IEEE TCC 2016
16. 3/13/2017 16Demetris Trihinas 16
P2P Monitoring
ā¢ Monitored elements form Gossip Overlay Network to propagate monitoring
info -> ease coordination and query computation
ā¢ Gossip denotes periodic and pairwise propagation of the current state between
monitored peers
āMonitoring Elastically Adaptive Multi-Cloud Servicesā, D. Trihinas, G. Pallis and M.D. Dikaiakos, IEEE TCC 2016
āSelf managing monitoring for highly elastic large scale cloud deploymentsā, S. Ward et al., ACM DIDC 2014
17. 3/13/2017 17Demetris Trihinas 17
Edge Computing
Init
Sense
Act
Sleep
Receive
Decide
Transmit
A term coined to reflect data processing and decision-making on
āsmartā devices that sit at the edge of IoT networks
Init
Sense
Transmit
Sleep
Interesting Articles
āThe promise of edge computingā, W. Shi and S. Dustdar, IEEE Computer, 2016
āThe swarm at the edge of the cloudā, E. Lee et al., IEEE Design & Test, Jun 2014
Simple
Sensing
Edge
Computing
<10ms
Big Data
Traffic
>100ms
ProcessingNo Processing
18. 3/13/2017 18Demetris Trihinas 18
Augmenting IoT with the Cloud
Challenge 1 ā Taming data volume and data velocity with limited
processing and network capabilities in the IoT/Edge realm
Challenge 2 - IoT devices are usually battery-powered which
means intense processing leads to less battery-life
Processing and data dissemination are
the main energy drains in embedded
and mobile devices
Interesting Article
āPractical data prediction for real-world wireless sensor networksā, U. Raza and T. Palpanas, IEEE TKDE, 2015
19. 3/13/2017 19Demetris Trihinas 19
Augmenting IoT with the Cloud
āTaking the internet to the next physical levelā, V. Cerf and M. Senges, IEEE Computer, 2016
Challenge 3 ā Security and data privacy
October 2016, DDOS Attack
21. 3/13/2017 21Demetris Trihinas 21
Preliminaries
ā¢ A monitoring stream š = {šš}š=0
š
published by a monitoring source is a large
stochastic sequence of i.i.d datapoints, denoted as šš , where š = 0, 1, ā¦ , š
and š ā ā
ā¢ A datapoint šš is a tuple ššš, š”š, š£š described by a unique identifier for the
monitoring source ššš, a timestamp š”š and a value š£š
ā¢ A datapoint may have a set of other attributes (e.g., location) although for
brevity we will omit other attributes without loss of generality
šš
šš+1
Monitoring Stream M
22. 3/13/2017 22Demetris Trihinas 22
Periodic Sampling
ā¢ The process of triggering the collection mechanism of a monitored source
every š time units such that the š š”ā datapoint is collected at time š”š = š ā š
šš
šš+1
Metric Stream M sampled every T = 1s
Compute resources and energy are wasted while generating large data volumes at a high velocity
Metric Stream M sampled every T = 10s
Sudden events and significant insights are missed
23. 3/13/2017 23Demetris Trihinas 23
More processing
(e.g., peak detection
for hr monitoring) Driving Peripherals
(e.g., accelerometer,
leds)
Interesting Articles
āEnergy-harvesting wearables for activity-aware servicesā, S. Khalifa et al., IEEE Internet Computing, 2015
āTaking the internet to the next physical levelā, V. Cerf and M. Senges, IEEE Computer, 2016
Emulated behavior of wearable device on Raspberry Pi with ARM processor
(1 core 32MHz, 128MB RAM)
Periodic Sampling with Physical
Sensing
OS Monitoring
24. 3/13/2017 24Demetris Trihinas 24
Periodic Sampling in an Energy Perspective
Lithium battery (35mAh) State of Charge (SoC) projection for a wearable
device with step, calorie and heartrate monitoring
Interesting Article
āA universal state-of-charge algorithm for batteriesā, B. Xiao et al., ACM DAC, 2010
25. 3/13/2017 25Demetris Trihinas 25
Periodic Dissemination
Monitoring
Source
Receiving
Entity
ā¢ The process of triggering the network controller of a monitored source every
š· time units such that the š š”ā
datapoint is disseminated at time š”š = š ā š· to
interested receiving entities
No data loss
ā¢ Aggregation-based policy dissemination:
ā¢ Withhold triggering dissemination until a window of š² datapoints is collected
ā¢ Withhold dissemination for fixed period of time and disseminate all datapoints
collected until then
26. 3/13/2017 26Demetris Trihinas 26
Periodic Dissemination Monitoring Cost
š š + š½š ā Ļ + Īµs
per message cost
per Ļ byte cost
datapoint d (t,v)
-> message compression
Interesting Article
āReal-Time Data Analytics in Sensor Networksā, T. Palpanas, ACM Sensors, 2013
energy for state transitions
Example Values (Mica2)
š š = .645 mJ
Īµs = .331 mJ
š½š = .0144 mJ/byte
28. 3/13/2017 28Demetris Trihinas 28
Preliminaries
ā¢ Approximate Techniques?
ā¢ Decision mechanisms based on estimation models capturing and predicting
monitoring stream runtime evolution within certain accuracy guarantees
ā¢ Adaptive?
ā¢ Adapt monitoring source properties based on the predicaments of the
estimation models
ā¢ Low-Cost?
ā¢ The costs (time, resources) of applying adaptive monitoring techniques are
(much) less than leaving the monitoring process as is
ā¢ We are interested in techniques running in-place and inexpensively
29. 3/13/2017 29Demetris Trihinas 29
Adaptive Sampling
ā¢ Dynamically adapt the sampling period šš based on some estimation model
š(š), containing runtime information of the monitoring stream evolution
ā¢ How large of an adjustment is required, āideallyā, depends on some evaluation
(distance) metric denoting the ability of the model to (correctly) estimate and
follow metric stream evolution
šš
šš+1
šš+1
Metric Stream šā² dynamically sampledMetric Stream š with š = 1š
30. 3/13/2017 30Demetris Trihinas 30
Adaptive Sampling
šš
šš+1
šš+1
Metric Stream šā² dynamically sampledMetric Stream š with š = 1š
Find the max š ā [š ššš, š ššš„] to collect šš+1 based on an estimation of
the monitoring stream evolution š(š), such that šā² differs from š less
than a user-defined imprecision value š¾ (ššš š” < š¾)
31. 3/13/2017 31Demetris Trihinas 31
Adaptive Dissemination
ā¢ Dynamically adapt dissemination rate by applying approximation
techniques to sensed datapoints to reduce communication overhead
ā¢ By suppressing consecutive datapoints from dissemination with
ālittleā change in their metric values
ā¢ How much is change is considered as ālittleā depends on:
ā¢ The approximation technique
ā¢ Accuracy guarantees given by the user and denoting the probability
(confidence Ī“) with which sensed datapoints are approximated
32. 3/13/2017 32Demetris Trihinas 32
Adaptive Dissemination (Model-Based)
ā¢ Monitoring source maintains runtime estimation model š(š) capturing monitoring
stream evolution and variability
ā¢ At the š š”ā
time interval instead of metric values, the model is disseminated
ā¢ Receiving entities predict the IoT device state from given model assuming subsequent
k-datapoints can be approximated di+k|i = f(š š , šš) within the given guarantees
Metric Stream š Approximated Metric Stream šā²
33. 3/13/2017 33Demetris Trihinas 33
Adaptive Dissemination (Model-Based)
ā¢ Monitoring source withholds further dissemination interacting with receiver
only when shifts in monitoring stream value distribution render model as
inconsistent with the actual IoT device state
ā¢ If model parameterization is inconsistent, at this point, it must be updated
decision criteria?
decision
function?
Metric Stream š Approximated Metric Stream šā²
34. 3/13/2017 34Demetris Trihinas 34
Metric Filtering
ā¢ A form of adaptive dissemination: suppress latest datapoint(s)
dissemination if their values have not āchangedā since last reported
ā¢ Metric stream suppression butā¦ no model -> no forecasting
ā¢ Fixed filter ranges assume monitoring stream value distribution will not
change in in the future (no guarantees any values will be filtered at all)
if (curValue ā [prevValue ā R, prevValue + R ])
filter(curValue)
Interesting Article
āMonitoring, aggregation and filtering for efficient management of virtual networksā, S. Clayman, IEEE NOMS, 2011
35. 3/13/2017 35Demetris Trihinas 35
Adaptive Filtering
ā¢ Dynamically adjust the filter range š based on the current variability of the
monitoring stream, denoted as š(š)
Metric Stream šā² with adaptive RangeMetric Stream š
šš
šš+1
š š+1
š š
ā¢ Find the max filter range š š+1 ā [š ššš, š ššš„] for šš+1 such that šā² differs
from š less than a user-defined imprecision value š¾ based on the variability
of the metric stream š(š)
37. 3/13/2017 37Demetris Trihinas 37
The Adaptive Monitoring Framework
Ī¼Processor
Ī¼OS
Memory
Low-Cost Approximate Stream Estimation
Adaptive
Sampling
Adaptive
Filtering
S
E
N
S
I
N
G
U
N
I
T
metric
updates
A
P
I
updated
periodicity
estimation
confidence
Adaptive
Dissemination
model updates &
buffer content
when shift
detected
updated filter range
stream variability
filtered metric stream
AdaM
N
E
T
W
O
R
K
U
N
I
T
Model
Base
stream
evolution
and online
stats
Datapoint
Buffer
IoT Device
metric
stream
compressed
metric
stream
38. 3/13/2017 38Demetris Trihinas 38
Low-Cost Approximate Estimation Model
ā¢ Three (3) phase approach to approximate monitoring stream
evolution and variability by exploiting knowledge (hidden) in
the monitoring stream
Step 1
Estimate Monitoring
Stream Evolution
and label datapoints
as ā(un)expectedā
Step 2
Detect if Gradual
Trends Exist in
Monitoring Stream
Evolution
Step 3
Test if Seasonality
Enrichment is
Beneficial to Stream
Evolution
39. 3/13/2017 39Demetris Trihinas 39
Low-Cost Approximate Estimation Model
Phase 1: Update runtime monitoring stream evolution š Ī
ā¢ Probabilistic Exponential Weighted Moving Average (PEWMA) to estimate
š£š+1 from šš and the standard deviation šš+1:
š£š+1 = šš = š ā ššā1 + 1 ā š š£š
ā¢ Datapoints labelled as āexpectedā if estimation lands in prediction intervals
determined from user confidence guarantees or āunexpectedā otherwise
ā¢ āUnexpectedā datapoints are stored in local buffer for dissemination
Looks like an exponential moving average, right?
But weighting š = š¼ 1 ā š½Pš is probabilistically applied!
40. 3/13/2017 40Demetris Trihinas 40
Low-Cost Approximate Estimation Model
EWMA after spikes
overestimates
subsequent values
EWMA slow to
acknowledge spikes
With probabilistic reasoning each datapoint will contribute to the
estimation process depending on itās p-value
Why use a Probabilistic EWMA?
41. 3/13/2017 41Demetris Trihinas 41
AdaM: Adaptive Sampling and Filtering
āAdaM: Adaptive Monitoring Framework for Sampling and Filtering on IoT Devicesā, D. Trihinas, G. Pallis and M.D. Dikaiakos, IEEE BigData 2015
ā[Low-Cost Adaptive Monitoring Techniques for the Internet of Thingsā, D. Trihinas, G. Pallis and M.D. Dikaiakos, IEEE TSC 2016
Ī¼Processor
Ī¼OS
Memory
Low-Cost Approximate Stream Estimation
Adaptive
Sampling
Adaptive
Filtering
S
E
N
S
I
N
G
U
N
I
T
metric
updates
A
P
I
updated
periodicity
estimation
confidence
Adaptive
Dissemination
model updates &
buffer content
when shift
detected
updated filter range
stream variability
filtered metric stream
AdaM
N
E
T
W
O
R
K
U
N
I
T
Model
Base
stream
evolution
and online
stats
Datapoint
Buffer
IoT Device
metric
stream
compressed
metric
stream
42. 3/13/2017 42Demetris Trihinas 42
Adaptive Sampling
ā¢ After updating estimation model, we compute current confidence (šš ā¤ 1)
of approach based on the actual and estimated standard deviation
ā¢ Next, we update sampling period šš+1 based on the current confidence
and the user-defined imprecision Ī³ (e.g. 10% tolerance to errors)
šš = 1 ā
| šš ā šš|
šš
ā¦ the more āconfidentā the algorithm is, the larger the
outputted sampling period šš+1 can beā¦
Ī» is an
aggressiveness
multiplier
(default Ī»=1)
43. 3/13/2017 43Demetris Trihinas 43
Adaptive Filtering
ā¢ Compute current variability of the metric stream using moving Fano
Factor š¹š which is an indicator of the stream value dispersion
ā¢ We then compare šššš to the user-provided maximum tolerable
imprecision guarantees š¾ in attempt to widen filter range
š¹š =
šš
2
šš
ā¦a low š¹š (due to šš) indicates a currently in-dispersed data
stream which means low variability in the metric streamā¦
45. 3/13/2017 45Demetris Trihinas 45
6x less processing! 4x less network traffic!
4x less energy usage!
+ AdaM
3 - 5 days
7 - 8 days
AdaM in action!
46. 3/13/2017 46Demetris Trihinas 46
Memory Trace
Java Sorting Benchmark
CPU Trace
Carnegie Mellon RainMon Project
Disk I/O Trace
Carnegie Mellon RainMon Project
TCP Port Monitoring Trace
Cyber Defence SANS Tech Institute
Step Trace
Fitbit Charge HR Wearable
Heartrate Trace
Fitbit Charge HR Wearable
More datasets!
47. 3/13/2017 47Demetris Trihinas 47
Balance between efficiency and accuracy!
Energy Consumption
CPU Cycles Network Traffic
Error (MAPE)
FAST [L. Fan et al., SIGMOD 2014] L-SIP [E Gaura et al., IEEE Trans. on Sensors, 2013]
48. 3/13/2017 48Demetris Trihinas 48
Dublin Smart City Intelligent Transportation Service (Dublin ITS)
.
.
.
.
.
.
1000 Buses* with GPS tracking
sending updates to ITS with 16
params (e.g. busID, location,
current route delay)
*Real data from Jan. 2014
Apache Kafka queuing
service on x-large VM
(16VCPU, 16GB RAM,
100GB Disk)
Apache Spark cluster with 5 workers on large VMs
(8VCPU, 8GB RAM, 40GB Disk)
Alerts ITS operators when more than 10 buses in a
Dublin city area, per 5min window, are reporting delays
over 1 standard deviation from their weekly average
AdaMā¦ integrated in big data streaming service!
49. 3/13/2017 49Demetris Trihinas 49
Spark total delay (processing + scheduling) for T=1, 5, 10 intervals and
AdaM with max imprecision Ī³ = 0.15
AdaMā¦ integrated in big data streaming service!
50. 3/13/2017 50Demetris Trihinas 50
AdaM achieves a >85% accuracy in major Dublin areas
AdaMā¦ integrated in big data streaming service!
51. 3/13/2017 51Demetris Trihinas 51
AdaM: Adaptive Dissemination
āADMin: Adaptive Monitoring Dissemination for the Internet of Thingsā, D. Trihinas, G. Pallis and M.D. Dikaiakos, IEEE INFOCOM, 2017
Ī¼Processor
Ī¼OS
Memory
Low-Cost Approximate Stream Estimation
Adaptive
Sampling
Adaptive
Filtering
S
E
N
S
I
N
G
U
N
I
T
metric
updates
A
P
I
updated
periodicity
estimation
confidence
Adaptive
Dissemination
model updates &
buffer content
when shift
detected
updated filter range
stream variability
filtered metric stream
AdaM
N
E
T
W
O
R
K
U
N
I
T
Model
Base
stream
evolution
and online
stats
Datapoint
Buffer
IoT Device
metric
stream
compressed
metric
stream
52. 3/13/2017 52Demetris Trihinas 52
AdaM: Adaptive Dissemination
āADMin: Adaptive Monitoring Dissemination for the Internet of Thingsā, D. Trihinas, G. Pallis and M.D. Dikaiakos, IEEE INFOCOM, 2017
Network
Unit
ADMin
Seasonality
Enrichment
Shift
Detection
A
P
I
Model
Base
Approximate
Stream
Estimation
Remote seasonal
periodicity
update
test if enrichment
is beneficial?
metric
updates
update
stream
evolution
and trend
seasonal
enrichment of
stream evolution
compressed
metric stream
Seasonal Periodicity
Detection
The ADMin
plugin for AdaM
Donāt disseminate
datapoint values...
Instead disseminate
model updates!
53. 3/13/2017 53Demetris Trihinas 53
Low-Cost Approximate Estimation Model
Step 2: Detect over time gradual trends in monitoring stream to reduce
ālaggingā effects in monitoring stream evolution estimation
š£š+š|š šš + š šš
Upward
trend
Downward
trend
ā¢ Holtās Trend Method used to bring moving average to
appropriate value base
ā¢ Improve forecasting from 1-step ahead (moving
average) predictions to k-datapoint values
54. 3/13/2017 54Demetris Trihinas 54
Low-Cost Approximate Estimation Model
Step 3: Test if seasonality enrichment is beneficial to estimation process
and update low-cost approximate model accordingly
seasonal
Seasonal with
exponential trend
Seasonal with
damped trend
ā¢ Tendency of the metric stream to exhibit behavior that
repeats itself every L periods (e.g., hourly)
ā¢ Seasonal effects highly evident in IoT data (e.g., human
biosignals, environmental data)
PV Panel Production Air Temperature
55. 3/13/2017 55Demetris Trihinas 55
Low-Cost Approximate Estimation Model
ā¢ Holt Winterās Method used to estimate seasonal contribution
ā¢ Forecasting k-subsequent datapoints with trend and seasonality
ā¢ Howeverā¦ perfect seasonal behavior is rarely observed in real-life
systems PV Panel Production
š£š+š|š šš + š šš + šš
Considering day before hourly
average (Si-L) will lead to
overestimation
56. 3/13/2017 56Demetris Trihinas 56
Low-Cost Approximate Estimation Model
ā¢ Online testing (T-Test) to determine if seasonal contribution beneficial
to estimation or not
ā¢ Detecting optimal seasonality cycle (L) is an open research challenge
especially when different cycles exist in monitoring stream
ā¢ Approximate runtime seasonal periodicity detection
ā¢ ComCube Framework (Matsubara et al., WWW, 2016): lightweight tensor-based
and parameter-free framework for near-optimal seasonal periodicity detection
š£š+š|š šš + š šš + šš
āNon-linear mining of competing local activities ā, Y. Matsubara, Y. Sakurai and C. Faloutsos, WWW 2016
57. 3/13/2017 57Demetris Trihinas 57
Detecting Shifts in a Monitoring Stream
ā¢ Cumulative Sum (CUSUM) to detect shifts in monitoring stream value
distribution which render estimation model as inconsistent
ts
prior shift
after shift
šš = ln
š(šš, šā²ā²
)
š šš, šā²
šš
ā²ā²
= šš
ā²
+ š
where Īµ magnitude
of change
1. log-likelihood ratio 2. Detecting shifts in mean
However, š not known beforehand
butā¦ estimation model model provides us with an approximate šŗ
58. 3/13/2017 58Demetris Trihinas 58
Detecting Shifts in a Monitoring Stream
3. Online CUSUM
4. Decision Function
Challenge 1: Linear timeā¦ Can šš be used
instead of š š?
ts ti
ti may greatly from ts differ in
cases of gradual trends
šš šŗš,{ššš¤,āššā} > ā ā š āššš” ššš”ššš”šš
h is measured
in standard
deviation units
5. Actual Shift Time
the time the
CUSUM detects
the shifts
59. 3/13/2017 59Demetris Trihinas 59
Detecting Shifts in a Monitoring Stream
ts ti
Trend and seasonality knowledge provide model with greater accuracy ( š ā š) and to
adapt to unexpected, abrupt and and volatile changes is monitoring stream
Challenge 2: CUSUM threshold h static and
sensitive when stream variability is low
(š ā 0 ) thus triggeringā¦ false alarms
Adapt CUSUM sensitivity by adapting
h after dissemination triggered and
restrict h with hmin
60. 3/13/2017 60Demetris Trihinas 60
ADMin Evaluation
ā¢ ADMin in 3 configurations
ā¢ No seasonality enrichment (ADMin)
ā¢ Static seasonality enrichment - previous day hourly average (ADMin_S1)
ā¢ Dynamic seasonality enrichment - ComCube integration (ADMin_S2)
ā¢ Under comparison frameworks
ā¢ LANCE [Werner et al., ACM SenSys 2011]
ā¢ G-SIP [Gaura et al., IEEE Trans. on Sensors 2013]
ā¢ ADWIN [Bifet et al., SIAM 2010]
All three framework parameters configured to output best results
61. 3/13/2017 61Demetris Trihinas 61
Photovoltaic Panel Current (IDC) Production
2 Weeks of data collected every 1 second
Data reduction: 87% -- Accuracy: 93%
Weather Station Air Temperature (oC)
2 Weeks of data collected every 1 second
Data reduction: 85% -- Accuracy: 92%
Wearable Human Heartrate (bpm)
2 Weeks of data collected every 1 minute
Data reduction: 80% -- Accuracy: 90%
ADMin in Action!
62. 3/13/2017 62Demetris Trihinas 62
Shift Detection Evaluation
ground truth
(offline PELT
algorithm)
false
alarms
correctly
detected
shifts
Air Temperature Dataset Wearable Dataset
Shift Detection Delay
63. 3/13/2017 63Demetris Trihinas 63
Energy Consumption
Data Volume Reduction
1O
interval
aggregation
ADWIN large energy consumption due to
complexity and absence of data reduction
scheme
LANCE large energy consumption due to enabling
dissemination x2 more than ADMin (false alarms)
and downloading each time all available datapoints
G-SIP large energy consumption due to shift
detection sensitivity (only EWMA based) which
enables false alarm disseminations
Overhead Evaluation
64. 3/13/2017 64Demetris Trihinas 64
Facebook DB Cluster Adaptive Monitoring
āInside the Social Networkās (Datacenter) Networkā, A. Roy et al., ACM SIGCOMM, 2015
65. 3/13/2017 65Demetris Trihinas 65
ā¢ Edge-mining on IoT devices is both resource and energy
intensive
ā¢ Big data streaming engines struggle to cope as the volume
and velocity of IoT-generated data keep increasing
ā¢ Adapts the monitoring intensity based on current metric
evolution and variability
ā¢ Reduces processing, network traffic and energy
consumption on IoT devices and the IoT network
ā¢ Achieves a balance between efficiency and accuracy
The AdaM Framework
Conclusion
67. 3/13/2017 67Demetris Trihinas 67
Low-Cost Approximate & Adaptive Monitoring
Techniques for the Internet of Things
Demetris Trihinas
trihinas@cs.ucy.ac.cy
PhD Candidate
Department of Computer Science
University of Cyprus
Notas do Editor
You cant have a gigantic pc on your wrist!
The building blocks of IoT
What makes IoT interesting and worth studying since we are obviously not device or hardware makersā¦ big data
So it only seems as natural that the cloud could be used
So it only seems as natural that the cloud could be used
Monitoring server (tree) hierarchy to increase fault-tolerance, robustness and scalability
Monitoring servers poll assigned agents that form monitoring clusters for metric updates pushing aggregated data further up the hierarchy
If you have large monitoring clusters there is a lot of network traffic going on intracluster-wise but if you have too small then the hierarchy grows too large and you have relaying costs when aggregating data and answering queries
These sensors are not source code sitting on VMsā¦. They are real devices scattered across the logical extremes of the network
(incoming/outcoming cloud traffic is priced)
Monitored elements (e.g., VMs, sensors)
These sensors are not source code sitting on VMsā¦. They are real devices scattered across the logical extremes of the network
These sensors are not source code sitting on VMsā¦. They are real devices scattered across the logical extremes of the network
How physical sensing differs from OS monitoringā¦ These sensors are not source code sitting on VMsā¦. They are real devices scattered across the logical extremes of the network
We use the word āideallyā because this is not the norm used by most techniques which use fixed step increments/decrements
So our confidence metric is actually a different way to denote this dist as we actually donāt have both the original and reconstructed version
Filtering (but its not really adaptive!)ā¦
Why probabilistic?
āUnexpectedā datapoints are stored in local buffer for dissemination (if needed e.g., for anomaly detection, otherwise as discussed only model updates need be sent to receivers)
Next slideā¦ datapoints contribute to estimation process based on p-value
Estimation base needs an uplift to not trail behind
Add faloutsos paper here
Shift in monitoring stream statistical propertiesā¦. We are not interested here for anomalies
Accumulating (small) portions of drift until threshold surpassed