Thermal Aware Workload Scheduling with Backfilling for Green Data Centers
Lizhe Wang, Gregor von Laszewski, Jai Dayal, Thomas R. Furlani
RIT . IU. UB
Outline
• Background and related work• Models• Research problem definition• Scheduling algorithm• Performance study• Conclusion
ContextCyberaide
A project that aims to make advanced
cyberinfrastructure easier to use
Future Grid
A newly NSF funded project to provide a
testbed that integrates the ability of dynamic
provisioning of resources.
(Geoffrey C. Fox is PI)
GreenIT & Cyberaide
How do we use advanced
cyberinfrastructure in an efficient way
GPGPU’s
Application use of special purpose
hardware as part of the cyberinfrastructure
FutureGrid• The goal of FutureGrid is to support the research that will
invent the future of distributed, grid, and cloud computing. • FutureGrid will build a robustly managed simulation
environment or testbed to support the development and early use in science of new technologies at all levels of the software stack: from networking to middleware to scientific applications.
• The environment will mimic TeraGrid and/or general parallel and distributed systems
• This test-bed will enable dramatic advances in science and engineering through collaborative evolution of science applications and related software.
University of Virginia (UV)Technical University DresdenGWT-TUD GmbH, GermanyUniversity of Tennessee – Knoxville (UTK)
Other Participant Sites
FutureGrid HardwareSystem type # CPUs # Cores TFLOPS RAM (GB)
Secondary storage (TB)
Default local file system Site
Dynamically configurable systems IBM iDataPlex 256 1024 11 3072 335* Lustre IU Dell PowerEdge 192 1152 12 1152 15 NFS TACC IBM iDataPlex 168 672 7 2016 120 GPFS UC IBM iDataPlex 168 672 7 2688 72 Lustre/PVFS UCSD Subtotal 784 3520 37 8928 542 Systems not dynamically configurable Cray XT5m 168 672 6 1344 335* Lustre IU Shared memory system TBD 40** 480** 4** 640** 335* Lustre IU Cell BE Cluster 4 IBM iDataPlex 64 256 2 768 5 NFS UF High Throughput Cluster 192 384 4 192 PU Subtotal 552 2080 21 3328 10 Total 1336 5600 58 10560 552
FutureGrid Partners• Indiana University• Purdue University• San Diego Supercomputer Center at University of California San
Diego• University of Chicago/Argonne National Labs• University of Florida• University of Southern California Information Sciences Institute,
University of Tennessee Knoxville• University of Texas at Austin/Texas Advanced Computing Center• University of Virginia• Center for Information Services and GWT-TUD from Technische
Universtität Dresden.
Green computing
• a study and practice of using computing resources in an efficient manner such that its impact on the environment is as less hazardous as possible.
– least amount of hazardous materials are used– computing resources are used efficiently in terms
of energy and to promote recyclability
Cyberaide Project• A middleware for Clusters, Grids and Clouds• A collaboration between IU, RIT, KIT, …• Project led by
Dr. Gregor von Laszewski
Objective
• Towards next generation cyberinfrastructure• Middleware for data centers, grids and clouds• Environment respect• To reduce temperatures of computing
resources in a data center, thus reduce cooling system cost and improve system reliability
• Methodology: thermal aware workload distribution
Model
• Data center– Node: <x,y,z>, ta, Temp(t)– TherMap: Temp(<x,y,z>,t)
• Workload– Job ={jobj}, jobj=(p,tarrive,tstart,treq,Δtemp(t))
t
RC-thermal model
Online task-temperature
Nodei.Temp(t)
Temp(Nodei.<x,y,z>,t)PR+
Nodei.Temp(0)
task-temperature profilenodei
<x,y,z>
ambient temperature:TherMap=Temp(Nodei.<x,y,z>,t)
Nodei.Temp(t)
P C R
Nodei.Temp(t)
Temp(Nodei.<x,y,z>,t)
Thermal model
Research issue definition
• Given a data center, workload, maximum temperature permitted of the data center
• Min Tresponse
• Min Temperature
Workload model
Data center model
TASA-B
Cooling system control
Workload placement
online task-temperature
input
schedule
input
input
Conceptframework
task-temperature profile
RC-thermal model
Workload model
Thermal map
Data center model
TASA-B
Cooling system control
Workload placement
calculation
online task-temperature
input
schedule
input
input
Conceptframework
task-temperature profile
RC-thermal model
Workload model
Thermal map
Data center model
TASA-B
Cooling system control
Workload placement
Control
calculation
online task-temperature
input
schedule
input
input
Conceptframework
task-temperature profile
RC-thermal model
Workload model
Thermal map
Data center model
TASA-B
Profiling tool
Cooling system control
Workload placement
Control
profiling
calculation
online task-temperature
input
schedule
input
input
Conceptframework
task-temperature profile
RC-thermal model
Workload model
Thermal map
Data center model
TASA-B
Profiling tool monitoring service
Cooling system control
Workload placement
Control
profiling
calculation
online task-temperature
CFD model
provide information Calculate thermal map
input
schedule
input
input
Conceptframework
Scheduling framework
Job submission
Jobs Job queue
Update data centerInformation periodically
Job scheduling Rack
Data center
TASA-B
Task scheduling algorithm with backfilling (TASA-B)
• Sort all jobs with decreased order of task-temperature profile
• Sort all resource with increased order of predicted temperature
• Hot jobs are allocated to cool resources• Predict resource temperature based on
online-task temperature• Backfill possible jobs
Node
Available time
t0
Time backfilling holes
nodek.tbfsta , backfilling start time of nodek
nodem
ax1
nodem
ax2
nodek.tbfend , end time for backfilling
Backfilling
nodem
ax1
Temperature
Tempbfmax
Node
Temperature backfilling holes
nodek.Tempbfsta, start temperature for backfilling of nodek
nodem
ax2
nodek.Tempbfend, end temperature for backfilling
Backfilling
Simulation
• Data center:– Computational Center for Research at UB – Dell x86 64 Linux cluster consisting 1056 nodes– 13 Tflop/s
• Workload:– 20 Feb 2009 – 22 Mar. 2009– 22385 jobs
Simulation resultMetrics TASAReduced average temperature 16.1 FReduced maximum temperature 6.1 FIncrease job response time 13.9%Saved power 5000 kWReduced CO2 emission 1900kg /hour
1 25 49 73 97 121 145 169 193 217 241 265 289 313 337 361 385 409 433 457 481 505 529 553 577 601 625 649 673 69770
80
90
100
110
FCFSTASA
Time (hour)
Aver
age
tem
pera
tue
(F)
Simulation resultMetrics TASA-BReduced average temperature 14.6 FReduced maximum temperature 4.1 FIncrease job response time 11%Saved power 4000 kWReduced CO2 emission 1600kg /hour
1 27 53 79 105131157183209235261287313339365391 417 443469495521 547573 599625651 6777030
20
40
60
80
100
120
FCFSTASA-B
Aver
age
tem
pera
ture
Our work on Green data center computing
• Power aware virtual machine scheduling (cluster’09)
• Power aware parallel task scheduling (submitted)
• TASA (i-SPAN’09)• TASA-B (ipccc’09) • ANN based temperature prediction and task
scheduling (submitted)
Final remark
• Green computing• Thermal aware data center computing• TASA-B• Justification with a simulation study