Transcript nxb-28-04-2009 - The Intelligent Modelling & Analysis

Introduction to Thermal Management
Naisan Benatar
Supervisors: Prof. Uwe Aickelin & Dr. Milena
Radenkovic
University of Nottingham
Outline
Introduction to WSNs
Problem Domain
Solution Outline
Future Work
IMA -University of Nottingham
2
Introduction – Who Am I?
Undergraduate Degree in CS – University of
Nottingham 2002-2005
Software Engineer – Thales Avionics – Worked
on SatCom system mainly for Airbus
Returned to UoN for PhD Studies in Sept
IMA -University of Nottingham
3
Areas of Research
Wireless Sensor Networks (WSNs)
Their Applications
Protocols related to their applications
IMA -University of Nottingham
4
What are Wireless Sensor
Networks?
•
Constructed from inexpensive nodes with a
sensor (or multiple sensors) and a wireless
networking device
•
•
Power is a concern – limited battery
Resource Constraints (CPU, Memory)
IMA -University of Nottingham
5
Applications of WSNs
Many Sizes and Applications:
Military
Conservation
Urban Monitoring (Traffic)
IMA -University of Nottingham
6
What I’m currently Interested In
Thermal Challenges in Data Centres –
Lots of Heat producing objects (mainly servers,storage etc )
Few cooling units (Active Air Conditioning Units)
Varying Loads mean varying temperatures
Each piece of equipment must not exceed its operating
conditions (Usually around 75 degrees Celsius)
Very heterogeneous environment – many manufacturers.
Changes in data centers are not uncommon (Often a 3 year
upgrade cycle)
IMA -University of Nottingham
7
IMA -University of Nottingham
8
Current Solution
Very Brute Force
All Coolers set so no piece of equipment goes
above a set level (approx 75 degrees Celsius)
Not very Intelligent
Wastes Energy
Does not adapt to varying workloads
No one system that handles all aspects of the thermal
environment
IMA -University of Nottingham
9
How can it improve?
Using a wireless sensor network composed of
numerous nodes equipped with temperature
sensor.
Gather Data from nodes
Make Decisions.
Be flexible, resilient and quick to respond or
predictive.
IMA -University of Nottingham
10
Problems with this approach
Lots of data – Possibly thousands of nodes in a
large DC.
Unordered/Unstructured data
Heterogeneous Environment
IMA -University of Nottingham
11
Potential Algorithm Inspiration
“Bio-inspired” Artificial Endocrine System
Traditional WSN approaches
IMA -University of Nottingham
12
Bio Inspired Approach
•
The human endocrine system regulates
processes in the body
•
E.g. Rate of breakdown of stored energy to
useable form controlled by 2 hormones (insulin &
glucagon)
•
Can something similar be used to regulate
cooling requirements in a DC?
IMA -University of Nottingham
13
Traditional Solutions
•
Directed Diffusion is a data centric protocol sometimes used
with WSNs.
•
Uses named data pairs, along with interests to specify
where data should be sent in the network.
•
Most useful for applications where all data ends up in a
single place for processing – not what we have here
IMA -University of Nottingham
14
How to test: A model
Need to build a Model (or 2) to simulate the various
algorithms:
Networking Model
Each node in the network will be individually modelled
Agent Based Modelling
->
Thermal Model
Thermal Environment will alter all nodes gradually
>System Dynamics?
IMA -University of Nottingham
-
15
Software
•
Using Anylogic (V 6.4)
•
•
•
Allows Combinations of Modelling Paradigms
Uses java for behaviour specification
Provides good foundation frameworks
IMA -University of Nottingham
16
How to get some confidence in the
model?
Performed experiments with real equipment –
measured temperature changes at varying levels
of load
Used as basis of thermal model
IMA -University of Nottingham
17
Measuring Performance
•
Metrics for Measurement of Performance of an
algorithm:
•
•
•
Packets sent
Time to respond to peaks
Energy Usage of cooling system
IMA -University of Nottingham
18
Real Life
Temps in Computer
60
50
Temp
40
Local Temp
CPU Temp
30
20
10
0
01/01/1900
01/01/1905
01/01/1910
01/01/1915
01/01/1920
Sample#
IMA -University of Nottingham
19
Simulation
Temp
70
60
50
40
Series1
30
20
10
0
1
9
17
25
33
41
49
57
65
73
81
89
97 105 113 121 129
IMA -University of Nottingham
20
Comparisons
•
•
Not very similar!
Many Reasons for this
•
•
•
Experimental Data not perfectly controlled
Many Simplifications in model
An ongoing topic in modelling research
IMA -University of Nottingham
21
Future Work – Short term
Improve Accuracy of model (CFD for thermal?).
Comparisons of different approaches to solving
the problem.
IMA -University of Nottingham
22
Future Work – Long term
Possible Experiments with small data centres
More Intelligence - predicative load based on
prior knowledge (e.g. Weekly peaks)
IMA -University of Nottingham
23
Questions?
IMA -University of Nottingham
24