Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal...
-
Upload
trinhnguyet -
Category
Documents
-
view
228 -
download
0
Transcript of Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal...
![Page 1: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/1.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
1
TA
CS
’04
Dynamic Thermal Managementfor
Distributed Systems
Andreas Weissel • Frank BellosaDepartment of Computer Science 4 (Operating Systems)
University of Erlangen-Nuremberg
Martensstr. 1, 91058 Erlangen, Germany
{weissel,bellosa}@cs.fau.de
![Page 2: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/2.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
2
TA
CS
’04
Benefits of Dynamic Thermal Management
■ Cooling servers, server clusters◆ cooling facilities often dimensioned for worst-case temperatures or
overprovisioned
■ Guarantee temperature limits◆ no need for overprovisioning of cooling units
◆ reduced costs (floor space, energy consumption, maintenance, ...)
■ Increased reliability◆ safe operation in case of cooling unit failure
◆ avoid local hot-spots in the server room
➜ Temperature sensors
![Page 3: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/3.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
3
TA
CS
’04
Drawbacks of Existing Approaches
■ If critical temperature is reached◆ throttle the CPU:
e.g. halt cycles, reduced duty cycle, reduced speed
■ But: neglect of application-, user- or service-specific requirementsdue to missing online information about◆ the originator of a specific hardware activation and
◆ the amount of energy consumed by that activity
➜ Throttling penalizes all tasks
![Page 4: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/4.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
4
TA
CS
’04
Outline
■ From events to energy◆ event-monitoring counters
◆ on-line estimation of energy consumption
■ From energy to temperature◆ temperature model
■ Energy Containers◆ accounting of energy consumption
◆ task-specific temperature management
■ Infrastructure for temperature management in distributed systems
![Page 5: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/5.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
5
TA
CS
’04
Approaches to Energy Characterization
■ Reading of thermal diode embedded in modern CPUs◆ low temporal resolution
◆ significant overhead
➜no information about originator of power consumption
![Page 6: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/6.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
6
TA
CS
’04
Approaches to Energy Characterization
■ Reading of thermal diode embedded in modern CPUs◆ low temporal resolution
◆ significant overhead
➜no information about originator of power consumption
■ Counting CPU cycles◆ time as an indicator for energy consumption
◆ time as an indicator for contribution to temperature level
◆ throttling according to runtime
➜but: wide variation of the active power consumption
![Page 7: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/7.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
7
TA
CS
’04
Approaches to Energy Characterization
■ P4 (2 GHz) running compute intensive tasks: CPU load of 100%◆ variation between 30–51 W
alu-
add
alu-
add-
bsw
ap-in
cal
u-m
ov-in
c-ad
dm
em-lo
adm
em-s
tore
push
pop
mem
-load
/sto
rem
em-in
c L1 L2ad
ler
ln2
sha
ripem
d160
wra
pfa
ctor
0
10
20
30
40
50
pow
er c
onsu
mpt
ion
[Wat
t]
idle powerconsumption:12.3 W
![Page 8: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/8.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
8
TA
CS
’04
From Events to Energy: Event-Monitoring Counters
■ Event counters register energy-critical events in the complete system architecture.◆ several events can be counted simultaneously
◆ low algorithmic overhead
◆ high temporal resolution
◆ fast response
■ Energy estimation
➜correlate a processor-internal event to an amount of energy
◆ select several events and use a linear combination of these event counts to compute the energy consumption
������ ����� ����⋅
∑=
![Page 9: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/9.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
9
TA
CS
’04
From Events to Energy: Methodology
■ Measure the energy consumption of training applications
■ Find the events with the highest correlation to energy consumption
■ Compute weights from linear combination of event counts and real power measurements of the CPU
➜solve linear optimization problem:find the linear combination of these events that produce the minimum estimation error
➜avoid underestimation of energy consumption
� ����� ���� ��������������–⋅
∑
��������������� � ����� ����⋅
∑≤
![Page 10: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/10.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
10
TA
CS
’04
From Events to Energy: Methodology
■ Set of events and their weights
■ Limitations of the Pentium 4◆ insufficient events for MMX, SSE & floating point instructions
◆ the case for dedicated Energy Monitoring Counters
event weight [nJ]
time stamp counter 6.17
unhalted cycles 7.12
µop queue writes 4.75
retired branches 0.56
mispred branches 340.46
mem retired 1.73
ld miss 1L retired 13.55
![Page 11: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/11.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
11
TA
CS
’04
Outline
■ From events to energy◆ event-monitoring counters
◆ on-line estimation of energy consumption
■ From energy to temperature◆ temperature model
■ Energy Containers◆ accounting of energy consumption
◆ task-specific temperature management
■ Infrastructure for temperature management in distributed systems
![Page 12: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/12.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
12
TA
CS
’04
From Energy to Temperature: Thermal Model
■ CPU and heat sink treated as a black box with energy in- and output
◆ energy input: electrical energy being consumed
◆ energy output: heat radiation and convection
CPUEnergy(#Events)
Thermal Capacity
Heat SinkThermal Resistance
Energy(∆Temp)
![Page 13: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/13.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
13
TA
CS
’04
From Energy to Temperature: Thermal Model
■ Energy input: energy consumed by the processor
CPUEnergy(#Events)
Thermal Capacity
Heat SinkThermal Resistance
Energy(∆Temp)
�� ����=
� �������������� ���
�
![Page 14: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/14.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
14
TA
CS
’04
From Energy to Temperature: Thermal Model
■ Energy output: primarily due to convection
CPUEnergy(#Events)
Thermal Capacity
Heat SinkThermal Resistance
Energy(∆Temp)
�� �� � ��–( )�–=
� � ���� �������
��
�� ����=
� �������������� ���
�
![Page 15: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/15.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
15
TA
CS
’04
From Energy to Temperature: Thermal Model
■ Altogether:
◆ energy estimator ➔ power consumption P
◆ time stamp counter ➔ time interval dt
◆ the constants , and have to be determined
�� ��� �� � ��–( )–[ ]�=
�� �� ��
![Page 16: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/16.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
16
TA
CS
’04
From Energy to Temperature: Thermal Model
■ Altogether:
◆ energy estimator ➔ power consumption P
◆ time stamp counter ➔ time interval dt
◆ the constants , and have to be determined
■ Solving this differential equation yields
�� ��� �� � ��–( )–[ ]�=
�� �� ��
� ( )��–��
------- ��– ����⋅ ���
����---- � ��+⋅+=
dynamic part static part
![Page 17: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/17.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
17
TA
CS
’04
Thermal Model: Dynamic Part
■ Measurements of the processor temperature◆ on a sudden constant power consumption and
◆ a sudden power reduction to HLT power.
➜fit an exponential function to the data: coefficient = ��
0 1000 2000 3000 4000 5000time [s]
30
35
40
45
50
55
60
tem
pera
ture
[° C
elsi
us]
IdleActiveIdle(50 W) (12 W)
Pentium 4@2GHz with active heat sink
![Page 18: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/18.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
18
TA
CS
’04
Thermal Model: Static Part
■ Static temperatures and power consumption of the test programs
10 20 30 40 50power consumption [Watt]
30
35
40
45
50
55
60st
atic
tem
pera
ture
[Cel
sius
]
![Page 19: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/19.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
19
TA
CS
’04
Thermal Model: Static Part
■ Linear function to determine and �� ��
10 20 30 40 50power consumption [Watt]
30
35
40
45
50
55
60st
atic
tem
pera
ture
[Cel
sius
]
![Page 20: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/20.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
20
TA
CS
’04
Thermal Model: Implementation
■ Linux 2.6 kernel
■ Periodically compute a temperature estimation from the estimated energy consumption
■ Deviation of a few degrees celsius over 24 hours◆ or if ambient temperature changes
■ Re-calibration with measured temperature every few minutes
![Page 21: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/21.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
21
TA
CS
’04
Thermal Model: Accuracy
0 2000 4000 6000 8000 10000 12000time [s]
35
40
45
50
55
60te
mpe
ratu
re (
cels
ius)
estimated temperaturemeasured temperature
![Page 22: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/22.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
22
TA
CS
’04
Outline
■ From events to energy◆ event-monitoring counters
◆ on-line estimation of energy consumption
■ From energy to temperature◆ temperature model
■ Energy Containers◆accounting of energy consumption◆ task-specific temperature management
■ Infrastructure for temperature management in distributed systems
![Page 23: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/23.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
23
TA
CS
’04
Properties of Energy Accounting
■ Accounting to different tasks/activities/clients◆ example: web server serving requests from different client classes
◆ e.g. Internet/Intranet, different service contracts
■ “Resource principal” can change dynamically
■ Client/server relationships between processes◆ account energy consumption of server to client
![Page 24: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/24.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
24
TA
CS
’04
Energy Containers
■ Resource Containers [OSDI ’99] ➔ Energy Containers◆ separation of protection domain and “resource principal”
■ Container Hierarchy◆ root container (whole system)
◆ processes are attached to containers
◆ this association can be changed dynamically (client/server relationship)
➜energy is automatically accounted to the activity responsible for it
■ Energy shares◆ amount of energy available (depending on energy limit)
◆ periodically refreshed
◆ if a container runs out of energy, its processes are stopped
web server
root
Intranet Internet
80% 20%
131.188.34.*
![Page 25: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/25.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
25
TA
CS
’04
Energy Containers
■ Example:web server working for two clients with different shares
0 100 200 300time (seconds)
0
10
20
30
40
50
60pow
er
consu
mptio
n (
Watt)
web server (total)client1 (80% share)client2 (20% share)
Start throttling
![Page 26: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/26.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
26
TA
CS
’04
Task-specific Temperature Management
■ Periodically compute an energy limit for the root container(depending on the temperature limit )
■ Dissolve to ➔
■ Energy budgets of all containers are limited according to their shares
■ Tasks are automatically throttled according to their contribution to the current temperature
■ Throttling is implemented by removing tasks from the runqueue
��
�� ��� ��+ � ��–( )[ ]��� ���� �–≤=
� ��
![Page 27: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/27.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
27
TA
CS
’04
Temperature Management
■ Example: Enforcing a temperature limit of 45º
0 500 1000 1500 2000 2500time [s]
35
40
45
50
55
tem
pera
ture
(ce
lsiu
s)
estimated temperaturemeasured temperature
temperature limit
![Page 28: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/28.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
28
TA
CS
’04
Outline
■ From events to energy◆ event-monitoring counters
◆ on-line estimation of energy consumption
■ From energy to temperature◆ temperature model
■ Energy containers◆ accounting of energy consumption
◆ task-specific temperature management
■ Infrastructure for temperature management in distributed systems
![Page 29: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/29.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
29
TA
CS
’04
Energy Containers
■ Distributed energy accounting
◆ Id transmitted with the network packets (IPv6 extension headers)
◆ receiving process attached to the corresponding energy container
◆ temperature and energy are cluster-wide accounted and limited
◆ transparent to applications and unmodified operating systems
id 1
root
Id 2 id 1
root
id 2 id 1
root
id 2
IPv6 IPv6
![Page 30: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/30.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
30
TA
CS
’04
Energy Containers
■ Energy accounting across machine boundaries◆ requests from two different clients represented by two containers
◆ web server sends requests to factorization server
➜the energy consumption of the server is correctly accounted to the client
0 500 1000
server #1 (web server)
time [s]0 500 10000
10
20
30
40
50
pow
er c
onsu
mpt
ion
[W]
server #2 (factorization)
time [s]1500 1500
container 1 (id 1)root container (total)
container 2 (id 2)
![Page 31: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/31.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
31
TA
CS
’04
Infrastructure for DTM in Distributed Systems
■ Distributed energy accounting
■ Foundation for policies managing energy and temperature in server clusters◆ account, monitor and limit energy consumption and temperature of
each node
■ Examples◆ set equal energy/temperature limits for all servers
➜cluster-wide uniform temperature and power densities, no hot spots in the server room
◆ use energy/temperature limits to
➜throttle affected servers in case of a cooling unit failure
➜reduce number of active cooling units in case of low utilization
![Page 32: Dynamic Thermal Management for Distributed Systemsskadron/tacs/weiss_slides.pdf · Dynamic Thermal Management for Distributed ... Dynamic Thermal Management for Distributed Systems](https://reader031.fdocuments.us/reader031/viewer/2022021900/5b595c997f8b9a657c8d2236/html5/thumbnails/32.jpg)
Dynamic Thermal Management for Distributed Systems© Andreas Weissel • University of Erlangen-Nuremberg • Computer Science 4 (Operating Systems) • 2004
32
TA
CS
’04
Conclusion
■ Event-monitoring counters enable◆ on-line energy accounting
◆ task-specific temperature management
■ Correctly account client/server relations across machine boundaries
■ Transparent to applications and unmodified operating systems
■ Future directions◆ examine more sophisticated energy models
◆ task-specific frequency scaling to adjust the thermal load