Download presentation
Presentation is loading. Please wait.
Published byMorgan Stephens Modified over 9 years ago
1
Accounting for Load Variation in Energy-Efficient Data Centers
Dzmitry Kliazovich University of Luxembourg, Luxembourg Pascal Bouvry Sisay T. Arzo University of Trento, Italy Fabrizio Granelli Samee U. Khan North Dakota State University
2
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Cloud Computing Cloud computing market: $241 billion in 2020 Main focus is on Software-as-a-Service (SaaS) Source: Larry Dignan, “Cloud computing market”, ZDNet, 2011. Jun 12, 2013 Dzmitry Kliazovich
3
Energy Efficiency of Data Centers
Server Load Server Hardware Jun 12, 2013 Dzmitry Kliazovich
4
Energy Consumption in Data Centers
For every $1 invested in hardware 80¢ are spent on energy Source: IDC, ‘Worldwide Server Power and Cooling Expense ’ Jun 12, 2013 Dzmitry Kliazovich
5
Load Variation in Data Centers
Demand for cloud applications varies Daily, weekly, monthly, … DC costs >$150M and takes 2+ years to design and build “Above the Clouds: A Berkeley View of Cloud Computing,” EECS Department UC Berkeley, 2009. Jun 12, 2013 Dzmitry Kliazovich
6
Load Variation in Data Centers
Typically over-provisioning Idle infrastructure Demand Capacity Time Resources Unused resources “Above the Clouds: A Berkeley View of Cloud Computing,” EECS Department UC Berkeley, 2009. Jun 12, 2013 Dzmitry Kliazovich
7
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Main Contributions Evaluating energy-efficient scheduler with job concentration policy Estimating the amount of hardware resources required to accommodate incoming jobs for the time sufficient to wake up additional hardware resources Evaluating possibility of using energy-efficient hardware Calculate economic effect of using energy-efficient hardware Jun 12, 2013 Dzmitry Kliazovich
8
Energy Consumption in Data Centers
Data centers consume 1.5% of all electricity consumed in the world, but only 15% efficient Jun 12, 2013 Dzmitry Kliazovich
9
Existing Power Management Solutions
Dynamic Power Management (DPM) Brings most of the savings as the average load often stays below 30% Dynamic Voltage and Frequency Scaling (DVFS) Adjust the hardware performance and power consumption to match the corresponding level of the load Jun 12, 2013 Dzmitry Kliazovich
10
Existing Power Management Solutions
Energy Model for Hosts memory modules, disks, I/O resources CPU Idle server consumes of the peak load for all CPU frequencies Jun 12, 2013 Dzmitry Kliazovich
11
Existing Power Management Solutions
Energy model for a network switch Chassis ~ 36% Linecards ~ 53% Port transceivers ~ 11% Jun 12, 2013 Dzmitry Kliazovich
12
Data Center Load Variation
Data centers are designed with over provisioning of computing and communication resources On average, datacenters are loaded only at 30% of their capacity In idle state, the servers consume around two thirds of its peak power consumption Jun 12, 2013 Dzmitry Kliazovich
13
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Load Distribution Jun 12, 2013 Dzmitry Kliazovich
14
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Data Center Hardware Blade Servers HP blade system c-Class, PowerEdge M610x Blade, and IBM BladeCenter HS23 Express 1027 W of energy and costs $2,323 on average Rack Servers HP DL380 , PowerEdge R420, and IBM System x3550 M3 225 W, but costs almost 79% more Jun 12, 2013 Dzmitry Kliazovich
15
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Data Center Hardware Energy-Efficient Servers Seamicro SM CPU per system 512 Intel Atom processor N570 1.66 GHz, 2 cores/4 threads 4 GB, DDR2 SODIMM Jun 12, 2013 Dzmitry Kliazovich
16
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Data Center Hardware Jun 12, 2013 Dzmitry Kliazovich
17
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Data Center Hardware Rack and blade servers increased energy consumptions Energy-efficient servers low power architectures, but more expensive, less computationally powerful Jun 12, 2013 Dzmitry Kliazovich
18
Cloud Computing Simulator
available at Measures cloud performance and energy efficiency First to simulate cloud communications with packet-level precision Implements network-aware scheduling Implements complete TCP/IP protocol stack Jun 12, 2013 Dzmitry Kliazovich
19
GreenCloud: Screenshots
Jun 12, 2013 Dzmitry Kliazovich
20
GreenCloud: Screenshots
Jun 12, 2013 Dzmitry Kliazovich
21
GreenCloud: Screenshots
Jun 12, 2013 Dzmitry Kliazovich
22
GreenCloud: Data Center Architectures
Supported data center architectures Two/Three-tier data centers Modular data centers Jun 12, 2013 Dzmitry Kliazovich
23
GreenCloud: Data Center Architectures
Future data center architectures FiConn DCell BCube Jun 12, 2013 Dzmitry Kliazovich
24
GreenCloud Architecture
Jun 12, 2013 Dzmitry Kliazovich
25
GreenCloud Usage and Benefits
GreenCloud tools cover complete optimization workflow Can be used to Optimize existing data centers Guide capacity extension decisions Help to design future data center facilities 1: Client data center analysis 2: Applying optimization solutions 3: Verification and proof of the concept Jun 12, 2013 Dzmitry Kliazovich
26
GreenCloud Usage and Benefits
Download at As source code As preinstalled virtual machine (VM) Jun 12, 2013 Dzmitry Kliazovich
27
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Simulation Scenario Datacenter architecture Three-tier fat tree Topology 1536 computing servers 32 racks served by 4 core and 8 aggregation switches Average datacenter load is 30% Simulation time 60 minutes Jun 12, 2013 Dzmitry Kliazovich
28
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Simulation Results Datacenter load variation Jun 12, 2013 Dzmitry Kliazovich
29
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Simulation Results Datacenter load distribution Jun 12, 2013 Dzmitry Kliazovich
30
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Simulation Results Change in the number of loaded servers Jun 12, 2013 Dzmitry Kliazovich
31
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Simulation Results Hardware and energy costs for partially loaded servers Jun 12, 2013 Dzmitry Kliazovich
32
Dzmitry Kliazovich (dzmitry.kliazovich@uni.lu)
Conclusions Energy-efficient hardware is costly, but can lead to significant cost benefits in data centers It is important to estimate the amount of hardware resources required to accommodate incoming jobs for the time sufficient for waking up additional hardware resources Design of scheduling policies should be guided by particularities in power management of the hardware components The obtained results confirm savings of up to $750 per server per year on average or a total of $300,000 in small data centers. Future work focuses on QoS aspects of job execution, use of job arrival traces from partner data centers, and prototype implementation of the designed methodology Jun 12, 2013 Dzmitry Kliazovich
33
University of Luxembourg
Thank you! Contact information: Dzmitry Kliazovich University of Luxembourg
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.