Towards Energy-Efficient Reactive Thermal Management in Instrumented - - PowerPoint PPT Presentation
Towards Energy-Efficient Reactive Thermal Management in Instrumented Datacenters Ivan Rodero 1 , Eun Kyung Lee 1 , Dario Pompili 1 , Manish Parashar 1 , Marc Gamell 2 , Renato J. Figueiredo 3 1 NSF Center for Autonomic Computing, Rutgers
Energy Efficient Grids, Clouds and Clusters, Brussels , October 26, 2010 Ivan Rodero1, Eun Kyung Lee1, Dario Pompili1, Manish Parashar1, Marc Gamell2, Renato J. Figueiredo3
1 NSF Center for Autonomic Computing, Rutgers University, NJ, USA 2 Open University of Catalonia, Barcelona, Spain 3 NSF Center for Autonomic Computing, University of Florida, FL, USA
Context and Motivation Datacenter Thermal Management Energy Efficiency and Tradeoffs Evaluation Methodology Results Next Steps Conclusions
2
3
Physical Environment
Sensor
Observer ¡
Actuator
Controller ¡
Resources
Sensor
Observer ¡
Actuator
Controller ¡
Virtualization
Sensor
Observer ¡
Actuator
Controller ¡
Application/ Workload
Sensor
Observer ¡
Actuator
Controller ¡
Cross-layer Power Management Cross-infrastructure Power Management Cloud
(private, public, hybrid, etc.)
Cloud
(private, public, hybrid, etc.)
Virtualized Instrumented infrastructure Application-aware
Goal: Autonomic (self-monitored and self-managed) computing systems:
Optimizing energy efficiency while ensuring Quality of Service delivered (performance)
4
Application/ Workload Virtualization Resources Physical Environment
Sensor 1 Sensor 2 Sensor 3 Sensor 4
Observer ¡1 ¡ Applica3on ¡req. ¡profiles ¡ Observer ¡2 ¡ VM ¡efficiency ¡ Observer ¡3 ¡ Resource ¡performance ¡ Observer ¡4 ¡ Environment ¡predic3on ¡ Local Controller Global Controller Observer ¡ Correla3ons ¡
Actuator Actuator 1 Actuator 2 Actuator 3
Local Controller Local Controller Local Controller Observer ¡ Controls ¡
Managed ¡Environment ¡ Actuator ¡ Sensor ¡ Controller’s ¡Command ¡ Observer’s ¡ ¡Sensing ¡Port ¡ Informa3on ¡Flow ¡ Request ¡Flow ¡ Resource ¡Flow ¡
Abnormal operational state detection
Distributed Online Clustering (e.g., workload) Physical sensing physical layer (e.g., thermal hotspots)
Reactive and proactive approaches
Reacting to anomalies to return to steady state Predict anomalies in order to avoid them
5
QoS Energy Efficiency Thermal efficiency
Abnormal state Steady State
QoS Energy Efficiency Thermal efficiency QoS Energy Efficiency Thermal efficiency
Different paths for reaching steady operational state Cross-layer actions (AUTONOMIC)
6
Workload Characterization (e.g., DOC) Pinning VM Migration Component-level Power Management
HPC Workload
Environment monitoring (temperature, etc.) Resources monitoring (load, power, etc.) Global controller Observer (correlations) Cooperate Proactive configuration Provisioning and maping Reactive configuration Reactive Designs for aggressive power management Depend Scheduling Trading with 3rd parties
8
20 40 60 5 10 15 2 2 4 6 8 Time [min] Node Number Temperature [C] 20 40 60 5 10 15 2 2 4 6 8 Time [min] Node Number Temperature [C]
Temporal correlation of the measured temperature under different workload distributions
120 140 160 180 200 220 240 260 200 400 600 800 1000 1200 1400 1600 1800 2000 Power (W) Time (s)
9
30 35 40 45 50 55 60 200 400 600 800 1000 1200 1400 1600 1800 2000 Temperature (C) Time (s) Internal Server Temperature Environment (Hotspot)
Correlation between server’s temperature and power
Reaction: VM migration
Steady
Assumption:
Reducing the activity factor (α)
VM Migration: move a VM to another server
May reduce CPU activity, but also memory activity, etc.
Overhead (suspend, transfer data, resume, etc.) Requires availability in another server (impact on the target
server)
10
cpu ≈ C × α × V 2 × f
Reducing the activity factor (α)
Pinning (in Xen platform): affinity in VCPUs – PCPUs mapping
CPUs without VMs running on them OS power management may result in automatic DVFS Penalty on the performance (resource sharing)
Reducing the frequency/voltage of CPUs ( V2×f )
Processor DVFS
Penalty on the performance (in general higher response time) Different possibilities
11
Goal: selection of appropriate technique to mitigate the
Energy-Efficient
Lower energy consumption Lower maximum/average power dissipation
Driven by optimization requirements. Examples:
Reduce temperature 5 oC (based on a threshold) A penalty of up to 10% on response time is acceptable
There are well known tradeoffs between performance and
energy efficiency
But also we need to consider other dimensions such as thermal
efficiency (temperature)
12
Example: Tradeoff between temperature and
13
Server configuration:
Two servers based on Intel quad-core Xeon processors
Operate at four frequencies ranging from 1.6GHz to 2.4GHz (but only 3
available under Xen 3.1)
CentOS Linux operating system with a patched 2.6.18 kernel with
Xen version 3.1
Additional hardware:
A “Watts Up? .NET” power meter to empirically measure
“instantaneous” power
Accuracy of ±1.5% of the measured power with sampling rate of 1Hz
TelosB motes to measure both internal (not sensors embedded into
the CPU) and external temperatures
A Sunbeam SFH111 heater (directed to the servers) in order to
emulate a thermal hotspot
Workload: HPL linpack
15 Use case: No running VMs in target server before migration
16
30 35 40 45 50 55 200 400 600 800 1000 1200 1400 1600 External Temperature Time (s) 4CPUs at 2.40GHz 2CPUs at 1.60GHz 4CPUs at 2.13GHz 4CPUs at 1.60GHz 30 32 34 36 38 40 42 44 46 200 400 600 800 1000 1200 1400 1600 External Temperature Time (s) 4CPUs at 2.40GHz 2CPUs at 1.60GHz 4CPUs at 2.13GHz 4CPUs at 1.60GHz 120 140 160 180 200 220 240 260 200 400 600 800 1000 1200 1400 1600 Power (W) Time (s) 4CPUs at 2.40GHz 2CPUs at 1.60GHz 4CPUs at 2.13GHz 4CPUs at 1.60GHz 30 32 34 36 38 40 42 44 46 200 400 600 800 1000 1200 1400 1600 External Temperature Time (s) Reference Migrate 1VM Migrate 2VMs Migrate 3VMs 30 35 40 45 50 55 200 400 600 800 1000 1200 1400 1600 External Temperature Time (s) Reference Migrate 1VM Migrate 2VMs Migrate 3VMs 120 140 160 180 200 220 240 260 200 400 600 800 1000 1200 1400 1600 Power (W) Time (s) Reference Migrate 1VM Migrate 2VMs Migrate 3VMs 30 35 40 45 50 55 200 400 600 800 1000 1200 1400 1600 Internal Temperature Time (s) Reference Pinning VMs to 3CPUs Pinning VMs to 2CPUs Pinning VMs to 1CPU 30 32 34 36 38 40 42 44 46 200 400 600 800 1000 1200 1400 1600 External Temperature Time (s) Reference Pinning VMs to 3CPUs Pinning VMs to 2CPUs Pinning VMs to 1CPU 120 140 160 180 200 220 240 260 200 400 600 800 1000 1200 1400 1600 Power (W) Time (s) Reference Pinning VMs to 3CPUs Pinning VMs to 2CPUs Pinning VMs to 1CPU
Internal Temperature oC Internal Temperature oC Internal Temperature oC External Temperature oC External Temperature oC External Temperature oC Power (W) Power (W) Power (W)
Correlation between internal and external temperature Correlation between temperature and power DVFS: using 2 CPUs at 1.6 GHz presents similar results than using 4 CPU at 2.13 GHz
17
Autonomic VM allocation and reactive technique decision
Cross-layer design approach
Examples: component-level power management, workload
clustering, etc.
Application-aware (workload characterization into CPU-, I/O-, etc.
intensive )
Optimization targets based on self-monitoring Models are required
VM migration, DVFS (work presented in this presentation) VM allocation (#VMs, workload characteristics, combinations,
etc.)
Models at the server and datacenter level
18
Tradeoffs exist between:
Performance Energy efficiency Thermal efficiency
Pinning is an effective mechanism to react to thermal anomalies
In addition to VM migration In contrast to DVFS
Different mechanisms’ behaviors observed depending on the
Autonomic decision making is required Cross-layer designs should improve datacenter’s management
20
21
http://nsfcac.rutgers.edu/GreenHPC/