Date post: | 12-Jan-2016 |
Category: |
Documents |
Upload: | liliana-mcdonald |
View: | 214 times |
Download: | 0 times |
Green HPC:Power Aware Scheduling of Bag-of-Tasks Applications with
Deadline Constraints on DVS-Enabled Data Centers
Kyong Hoon Kim1, Rajkumar Buyya1, and Jong Kim2
1Grid Computing and Distributed Systems (GRIDS) LaboratoryDept. of Computer Science and Software EngineeringThe University of Melbourne, Australiawww.gridbus.org
2POSTECH, Korea
Gridbus Sponsors
2
Outline
Introduction Related Work System Model Job Admission Control DVS-based Cluster Scheduling
EDF-based scheduling Proportional share-based scheduling
Simulation Results Summary
3
Background
Traditionally, high-performance computing (HPC) community has focused on performance (speed).
At the same time microprocessor vendors have not only doubled the number of transistors (and speed) every 18-24 months, but they have also doubled the power densities.
Moore’s Law for Power Consumption:
4
Research Motivations of Power Aware/Energy Efficient High Performance
Computing (HPC) Rapid uptake of HPC-architecture based Data Centers for
hosting industrial applications Reducing the operational costs of powering and cooling HPC
systems The tremendous increase in computer performance has come with an even
grater increase in power usage. According to Eric Schmit, CEO of Google, what matter most to Google is
“not speed but power, because data centers can consume as much electricity as a city.”
Improving reliability As a rule of thumb, for every 10°C increase in temperature, the failure rate
of a system doubles. Computing environment affected the correctness of the results.
The 18-node Linux cluster produced an answer outside the residual (i.e., a silent error) when running in dusty 85°F warehouse but produced the correct answer when running in a 65°F machine-cooled room.
5
Reliability/Implications
Reliability of Leading Edge Supercomputer (D. Reed, 2004)
Estimated Cost of An hour of system downtime (W. Feng, (ACM Queue, 2003):
6
Power Aware Computing
Power Aware (PA) computing/communications The objective of PA computing/communications is to improve power
management and consumption using the awareness of power consumption of devices.
Power consumption is one of the most important considerations in mobile devices due to the limitation of the battery life.
System level power management Recent devices (CPU, disk, communication links, etc.) support multiple
power modes. System scheduler can use these multiple power modes to reduce the
power consumption.
7
Outline
Introduction Related Work System Model Job Admission Control DVS-based Cluster Scheduling
EDF-based scheduling Proportional share-based scheduling
Simulation Results Summary
8
Related Work (1/3)
Research on power reduction for scientific applications
Hsu and Feng (SC 2005) – Los Alamos National Lab, USA -adaptation algorithm Automatic adaptation of CPU frequencies : the intensity level of off-chip accesses
Ge, Feng, and Cameron (2005) Three DVS scheduling strategies Software framework to implement scheduling techniques
Hotta, et. al. (2006) Profile-based power-performance optimization Selection of an appropriate gear using DVS scheduling Development of power-profiling system called PowerWatch
…
9
Related Work (2/3)
Energy reduction for MPI programs Kappiah, et. al. (2005) - NC State, and Georgia Uni, USA
Inter-node bottle problem in MPI programs Selection of an appropriate gear based on slack time
Lim, et. al. (2006) – NC State, and Georgia Uni, USA Adaptive DVS of Communication Phases in MPI programs
Son, et. al. (2006) …
Two approaches for building power-aware cluster platforms
Design and develop systems with consideration of energy consumption.
BlueGene/L, Green Destiny, … Use DVS-enabled commodity systems.
Clusters with AMD Athlon64s, Pentium Ms, AMD Opterons, …
10
Related Work (3/3)
DVS (Dynamic Voltage Scaling) technique Reducing the dynamic energy consumption by lowering the supply voltage at
the cost of performance degradation Recent processors support such ability to adjust the supply voltage
dynamically. The dynamic energy consumption = * Vdd2 * Ncycle
Vdd : the supply voltage Ncycle : the number of clock cycle
An example
10 msec 25 msec
deadlinePower
5.02
(a) Supply voltage = 5.0 V
10 msec 25 msec
deadlinePower
2.02
(b) Supply voltage = 2.0 V
11
DVS-based Power Aware Cluster Scheduling
Research motivation Previous work has focused on the development of DVS-
enabled cluster systems. Few works have considered the scheduling problem in power-
aware clusters.
Problem to solve To provide scheduling algorithms in DVS-enabled cluster
systems in order to minimize the energy consumption and to meet the job deadline.
Exploit industries move towards Utility Model / SLA-based Resource Allocation…
12
Outline
Introduction Related Work System Model Job Admission Control DVS-based Cluster Scheduling
EDF-based scheduling Proportional share-based scheduling
Simulation Results Summary
13
System Model (1/2)
Cluster model A cluster system is defined as (N, Q).
N: the number of processors Q: the processing performance of each PE in terms of MIPS
Job model A job is considered to be a bag-of-tasks application. The deadline is used as a QoS parameter of a job. A job = (p, {l1, l2, …, lp}, d)
p : the number of sub-tasks li : the length in MI of the i-th task d: the job deadline
14
System Model (2/2)
Energy model Energy consumption of a task execution
E = V2L L : the task length V : the supply voltage : a proportional constant
Dynamic Voltage Scaling {V1, …, Vm} : m different voltage levels Qi : the processor speed (MIPS) under the associated voltage level Vi Si : the normalized speed of each voltage level Vi (Si = Qi/Qm)
An Example
Voltage (Vi) MIPS (Qi) Relative Speed (Si)0.9 V 4,000 0.41.1 V 6,000 0.61.3 V 8,000 0.81.5 V 10,000 1.0
15
Outline
Introduction Related Work System Model Job Admission Control DVS-based Cluster Scheduling
EDF-based scheduling Proportional share-based scheduling
Simulation Results Summary
16
Proposed Cluster RMS System Architecture with Energy-Efficient
Resource Allocation
(1) Job submission (2) Schedulability test & Energy estimation (3) Acknowledgement of schedulability and energy amount (4) Selection of PEs
ResourceController
Processor
EnergyEstimator
Job-Queue
PE1
Processor
EnergyEstimator
Job-Queue
PE2
Processor
EnergyEstimator
Job-Queue
PEN
User
User
User
(1)(2)
(3)
(4)
Cluster
17
Application Admission and Resource Allocation Algorithm
Can task be completed in this PE?
How to estimate energy consumption?
We propose two DVS scheduling algorithms based on • Space-shared policy (EDF)• Time-shared policy
Algorithm Admission_Resource_Allocation (J = (p, {l1, …, lp}, d)) 1: for i from 1 to p do 2: PEalloc null; 3: energymin MAX_VALUE; 4: for k from 1 to N do 5: if schedulable (PEk, li, d) == true then 6: energyk energy_estimate (PEk, li, d); 7: if energyk < energymin then 8: energymin energyk; 9: PEalloc PEk;10: endif11: endif12: endfor13: if PEalloc != null then14: Allocate the i-th task of J to PEalloc;15: else16: Cancel all tasks of J.17: return reject;18: endelse19: endfor20: return accept;
18
Outline
Introduction Related Work System Model Job Admission Control DVS-based Cluster Scheduling
EDF-based scheduling Proportional share-based scheduling
Simulation Results Summary
19
EDF-based on DVS scheduling (1/4)
Basics Tk = {k,i (ek,i, dk,i) | i = 1, …, nk}
The current available task set in the k-th PE nk : the current number of tasks
k,i (ek,i, dk,i) : the i-th task in Tk
ek,i : the remaining execution time
dk,i : the remaining deadline
EDF (Early Deadline First) policy Tk is sorted by the deadline so that dk,i dk,i+1
The scheduler always executes the earliest-deadline task in the queue.
20
EDF-based DVS scheduling (2/4)
The temporary utilization, uk,i
The required processor utilization for task k,i by EDF
The continuous speed level of the highest-priority task, sk
The supply voltage level of the highest-priority task, vk
ik
i
j jk
ik d
eu
,
1 ,
,
}{max~,1 ik
nik us k
}~|{min 1 kiimik sSVv
~
21
EDF-based DVS scheduling (3/4)
An example Tk = {k,1(1, 4), k,2(2, 6), k,3(2, 10)} Temporary utilizations at time 0
uk,1 = 1/4 uk,2 = (1 + 2)/6 = 1/2 uk,3 = (1 + 2 + 2)/10 = 1/2
Scaling factors At time 0
sk = max{uk,1, uk,2, uk,3} = 1/2 vk = 1.1V
Energy model to use
0.60.4 k,1
5 10
Speed level
0
k,2 k,3
vk = 1.1V 1.1V10/6
0.9V
~
Voltage (Vi)
MIPS (Qi)Relative Speed
(Si)0.9 V 4,000 0.41.1 V 6,000 0.61.3 V 8,000 0.81.5 V 10,000 1.0
22
EDF-based DVS scheduling (4/4)
Schedulability test of EDF Energy estimation of EDF
Algorithm schedulable_EDF (PEk, l, d)Tk’ Tk {(l/Qm, d)};Sort Tk’ in the order of deadline.for i from 1 to nk + 1 do uk’,i ek’,i / dk,i ; if uk’,i > 1 then return false;endforreturn true;
Algorithm energy_estimate_EDF (PEk, l, d)Ecurrent energy_consumption (Tk, nk);Tk’ Tk {(l/Qm, d)};Enew energy_consumption (Tk’, nk+1);return (Enew – Ecurrent);function energy_consumption (T, n)Energy 0;time the current time;for i from 1 to n do for j from i to n do uj ek /dj; s’ max {uj} v min {Vj | Sj s’} s min {Sj | Sj s’} Energy Energy + v2eiQm; time time + ei/s; for j from i to n do dj dj – ei/s;endforreturn Energy;
23
Proportional Share-based DVS scheduling (1/2)
The proportional share scheme Multiple tasks share the processor performance in proportion to each task’s
weight. Each task should be given at least ek,i/dk,i under the maximum processor
speed to meet the deadline. The continuous processor speed level, sk
The supply voltage level of the highest-priority task, vk
The proportional share of each task, sharek
61 8 15
kn
i ikikk des1 ,, /~
k
k
n
j ikik
ikikn
j jk
jkk
ik
ik
kik
de
de
d
es
d
e
sshare
1 ,,
,,
1 ,
,
,
,,
/
/)(
1
}~|{min 1 kiimik sSVv
~
24
Proportional Share-based DVS scheduling (2/2)
An example Tk = {k,1(1, 4), k,2(2, 6), k,3(2, 10)}
Schedulability test and energy estimation is similar to EDF algorithm.
8 15
0.60.4
k,1 : 0.32
Speed level
0
vk = 1.3V 1.1V
3.906
0.9V
0.8
k,2 : 0.425
k,3 : 0.255
0 5.712
k,2 : 0.62
k,3 : 0.38
7.69
k,3 : 1.0
k,i : sharek,i
25
Outline
Introduction Related Work System Model Job Admission Control DVS-based Cluster Scheduling
EDF-based scheduling Proportional share-based scheduling
Simulation Results Summary
26
Simulation Environment
Using the GridSim toolkit A cluster system with 32 DVS-enabled processors
Operating points of simulated processors based on Athlon-64
1000 bag-of-tasks applications Task characteristics
Task length : 600,000 MIs ~ 7,200,000 MIs The number of tasks : 2 ~ 32 Deadline : 20% ~ 100% more than average execution time
Frequency Voltage MIPS Relative Speed
0.8 GHz 0.9 V 4,000 0.41.0 GHz 1.0 V 5,000 0.51.2 GHz 1.1 V 6,000 0.61.4 GHz 1.2 V 7,000 0.71.6 GHz 1.3 V 8,000 0.81.8 GHz 1.4 V 9,000 0.92.0 GHz 1.5 V 10,000 1.0
27
Simulated algorithms
DVS-based scheduling EDF-DVS PShare-DVS
Scheduling at maximum processor speed EDF-1.5V PShare-1.5V
Scheduling at minimum processor speed EDF-0.9V PShare-1.5V
28
Job Acceptance Rate
(a) EDF
0
20
40
60
80
100
2 3 4 5 6 7 8
Inter- arrival time (mins)
Acc
epta
nce
rate
(%
)
EDF- 1.5VEDF- DVSEDF- 0.9V
(b) PShare
0
20
40
60
80
100
2 3 4 5 6 7 8
Inter- arrival time (mins)
Acc
epta
nce
rate
(%
)
PShare- 1.5VPShare- DVSPShare- 0.9v
0
20
40
60
80
100
2 3 4 5 6 7 8
inter- arrival time (min)
EDF- DVSEDF- 0.9VEDF- 1.5VPShare- DVSPShare- 0.9VPShare- 1.5V
Acceptance rate (%)
29
Energy consumption
Normalized to EDF-1.5V at inter-arrival time of 2 mins.
Energy consumpation per accepted task
0
0.2
0.4
0.6
0.8
1
1.2
2 3 4 5 6 7 8
EDF- 1.5VPShare- 1.5VEDF- DVSPShare- DVSEDF- 0.9VPShare- 0.9V
Inter-arrival time (min)
Normalizedvalue
30
Normalized performance of DVS
Inter-arrival time (min)
EDF-DVS vsEDF-1.5V
PShare-DVS vsPShare-1.5V
Energy Reduction
(%)
Acceptance Degradatio
n (%)
Energy Reduction
(%)
Acceptance Degradatio
n (%)
2 13.6 13.3 33.8 14.3
3 21.3 13.0 34.4 12.8
4 31.2 11.2 36.3 9.7
5 34.4 7.9 38.8 9.5
6 38.6 6.3 42.8 7.0
7 41.5 4.0 43.7 5.6
8 44.3 2.7 45.2 5.2
31
Impact of granularity/number of controllable voltage levels
Normalized performance of EDF Normalized performance of PShare
0
0.2
0.4
0.6
0.8
1
1 2 3 4 7 13
Number of voltage levels
Normalized acceptance ratio
Normalized energy consumption
0
0.2
0.4
0.6
0.8
1
1 2 3 4 7 13
Number of voltage levels
Normalized acceptance ratio
Normalized energy consumption
Normalizedvalue
Normalizedvalue
32
Outline
Introduction Related Work System Model Job Admission Control DVS-based Cluster Scheduling
EDF-based scheduling Proportional share-based scheduling
Simulation Results Summary
33
Summary
Two primary drivers for Power-Aware HPC Operational cost Reliability
Power-aware scheduling with deadline constraints Reducing energy consumption Meeting jobs’ deadlines
The proposed scheduling algorithms DVS-based scheduling based on
Space-shared policy EDF Time-share policy / Proportional Resource Sharing
Minimizing cost under the constraint of the job deadline Future work
Budget-constrained power-aware scheduling Power-aware workflow scheduling
34
Thanks for your attention!
We Welcome Cooperation in Research and Development!http:/www.gridbus.org
eScience2007.org