Computing clusters are evaluated using different performance metrics, which often appear to be conflicting while being attempted to be optimized. For such conflicting cases along with frequently having an existence of heterogeneous environment, it is difficult for the cluster administrators to efficiently schedule machines, i.e., to select the right number and right combination of machines. In this paper, we develop a technique through which cluster administrators can select the right set of machines to enhance energy efficiency and cluster performance. To do so, first, we perform extensive laboratory experiments for a period of more than one year. Based on empirical analyses of data collected from the experiments, we formulate a many-objective optimization problem for clusters and integrate a greedy approach with Non-dominated Sorting Genetic Algorithm (NSGA-III) to solve this problem. We demonstrate that our approach mostly performs better than existing approaches in the literature through both real experimentation and simulation.
Presentation by Andreas Schleicher Tackling the School Absenteeism Crisis 30 ...
Exploiting a Synergy between Greedy Approach and NSGA for Scheduling in Computing Clusters
1. Exploiting a Synergy between Greedy
Approach and NSGA for Scheduling in
Computing Clusters
A S M Rizvi1, Tarik Reza Toha2, Siddhartha Shankar Das3, Sriram Chellappan4, A. B. M. Alim Al Islam5
NSysS 2018
Dhaka, Bangladesh
1
2, 3, 5 4
2. Overview of This Presentation
• Background and motivation
• Proposed methodology for clusters
– Many-objective problem formulation
– Solution method using NSGA-III and greedy approach
• Real experiment and simulation
– Performance evaluation of proposed methodology
• Conclusion
2
3. Background of Our Study
3
Galaxy formation Planetary movements Climate changes
Modeling, simulation, and experimentation of complex real-world
phenomena demand rigorous computing
Traffic simulation Plate tectonics Weather forecasts
4. Cluster topology
Grid computingCloud computing
Architectures for Parallel Computing
4
Large Linux cluster in University
of Technology, Germany
Cluster machines are connected
by a local area network
Clouds and grids are
geographically distributed
Clusters are tightly coupled,
whereas clouds and grids are
loosely coupled
Home-made cluster
5. Motivation Behind Our Study
Clusters have conflicting outcomes
Increasing the number of
machines
Decreases
computation time
Increases
maintenance cost
Desirable Not desirable
5
Optimizations required for:
1. # of machines in the cluster
2. Machines to be selected
6. Related Research Work
6
• Particle Swarm Optimization (PSO) based approach
– C. Lijun et al., IJGDC, 2015
– Utilizes a variant of PSO
– Focuses on load balancing from the perspectives of
• Average workload of all servers in cloud clusters
• Deviation of the workload
• Migration cost between servers
Does not integrate any empirical performance characterization of clusters
Does not accumulate the impact of cooling energy consumption
7. Related Research Work (cont.)
7
• Multi-objective optimization technique based on
Ant Colony Optimization (ACO)
– Y. Gao et al., Journal of Computer and System
Sciences, 2013
– Simultaneously minimizes two different objectives in
cloud architectures
• Resource wastage
• Power consumption
• Load balancing
Does not integrate any empirical performance characterization of clusters
Does not accumulate the impact of cooling energy consumption
8. Issues That Are Yet to Be Explored
8
•Consideration of cooling energy consumption
Source: http://www.ha-inc.com/blog/entry/data-center-energy-consumption/
Data center cooling energy consumption can depend on several factors like room size
to cooling capacity ratio, environmental temperature, design issue, etc. A good design
can reduce the total energy consumption.
•Empirical performance characterization of clusters
Should result in a new optimization model
9. Our Contributions
•We exploit a synergy between greedy method and
NSGA-III algorithm to solve a many-objective
optimization problem for clusters
̵ Incorporating cooling energy consumption
̵ Utilizing empirical characterization of clusters
•We perform rigorous experimentation to
demonstrate efficacy of our proposed solution
9
10. Formulation of Our Many-Objective
Optimization Problem for Computing Clusters
10
Following the empirical analysis,
we formulate our objective
functions as follows:
Computation
time
Energy
consumption
Cost
Inverse of
resource
utilization
Restriction on
assigned
workload
Constraint on
# of selected
machines
Limit of
cooling
temperature
11. Modification in The Existing NSGA-III
Algorithm: Modified Selection Process
11
Sharp change at
the beginning
providing results
below expectation
# of selected machines
should be greater than
a threshold. We select
the value as 6.
12. Modification in The Existing NSGA-III
Algorithm: Modified Crossover
12
Half Uniform Crossover
(HUX)
Greedy clustering
approach
Greedy approach to
include the best machine
and exclude the worst
machine
13. Modified Crossover: Half Uniform Crossover
13
Half-uniform crossover
Machine
selection
decision variables
(binary type)
Cluster
temperature
decision variable
(float type)
Parent 1:
Parent 2:
Yellow variables will have crossover within yellow variables
14. Modified Crossover: Greedy Clustering
Approach
Processor
speed
Memory
Network
B/W
Best
group
Worst
group
Best
group
Worst
group
Best
group
Worst
group
Find the
common
machines
Find the
common
machines
Biased to
include
machines
from this
group
Biased to
exclude
machines
from this
group
14
15. Modified Crossover: Greedy Approach to Include The
Best Machine and Exclude The Worst Machine
Processor
speed
Memory
Network
B/W
Best
machine
Worst
machine
Best
machine
Worst
machine
Randomly
select any one
property
Based on a
probability,
include the best
machine and
exclude the
worst machine
15
Best
machine
Worst
machine
16. Modification in The Existing NSGA-III
Algorithm: Solution Filtering
16
Each line indicates four objective values of modified NSGA-III
Worst objective
values for Objective
1, 3, and 4.
Best objective
value for Objective
2 only.
𝑭 𝒕𝒐𝒕𝒂𝒍 = 𝑾 𝒐𝒃𝒋𝟏 × 𝑽 𝒐𝒃𝒋𝟏 + 𝑾 𝒐𝒃𝒋𝟐 × 𝑽 𝒐𝒃𝒋𝟐 + … + 𝑾 𝒐𝒃𝒋𝑵 × 𝑽 𝒐𝒃𝒋𝑵
Pareto front showing 15 solutions of
our minimization problem
We avoid this kind of
solutions
17. Performance Evaluation of Our Proposed
Solution
• Real experiment over a testbed
– In a laboratory having 30 machines under a custom-built
cluster
• SimGrid simulation
– To enable experiments with more number of machines
17
Energy monitoring module for measuring
computation power and cooling power Real testbed setup
18. Experimental Setup for Comparing Modified NSGA-
III, PSO, and ACO in Real Testbed
18
Parameter Value
# of master machine 1
# of slave machines 29
Processor Intel Core 2 Duo
Processor base frequency 2.4 GHz, 2.66 GHz, and 2.8 GHz
Memory 1GB to 2GB
OS Ubuntu 14.04 LTS (x86)
Network B/W 5 to 100 MBPS
Total # of files 24, 16, and 11
Size of each file 787 MB
Total data size 18.9 GB, 12.6 GB, and 8.7 GB
Maximum allowable temperature 20℃-23℃
Environment temperature 28℃
20. Experimental Setup for Comparing NSGA-III,
PSO, and ACO in SimGrid
20
Parameter Value
# of master machines 1
# of slave machines 29, 49
Processor Intel Core 2 Duo
PC power Peak: 10 - 400 W, idle: 2.5-100 W,
power off: 5 W
Network B/W 10 to 100 kbps
Total # of files 86, 64, and 36
Size of each file 787 MB
Total data size 67.7 GB, 50.4 GB, and 28.3 GB
Maximum allowable temperature 20℃-23℃
Environment temperature 28℃
SimGrid version 3.12
25. Conclusion and Future Work
• Cluster administrator needs to decide the right number and
right combination of machines, where a many-objective
approach is yet to be focused in the literature
• We provide a solution for this problem considering cooling
energy consumption and empirical performance
characterization
• We formulate a synergy between NSGA-III and greedy
approach to solve this problem
• Future work
Analysis with machine failures
Analysis on different factors for cooling energy including room
size or cooling capacity
25
28. Cluster topology
Grid computingCloud computing
Architectures for Parallel Computing
28
Large Linux cluster in University
of Technology, Germany
Cluster machines are connected
by a local area network
Clouds and grids are
geographically distributed
Clusters are tightly coupled,
whereas clouds and grids are
loosely coupled
Home-made cluster
29. Background for Parallel Computing
29
Problem
Processor
The whole problem is distributed in small parts
to a single processor
Problem_1
Problem_2
Problem_3
Processor_1
Processor_2
Processor_3
Serial computing
Parallel computing
30. Related Research Work (cont.)
30
• Multi-objective optimization for virtual machine
based schemes in cloud
– R. Li et al., IEEE CLOUD, 2016
– Uses Biogeography Based Optimization (BBO)
technique for solving multi-objective optimization
– Considers only load balancing through resource
utilization in VMs placed in cloud architecture
Does not integrate any empirical performance characterization of clusters
Does not accumulate the impact of cooling energy consumption
31. Related Research Work (cont.)
31
• Stochastic optimization approach
– K.M. Tarplee et al., IEEE Transactions on Cloud
Computing (TCC), 2016
– Utilizes multi-objective stochastic programming
– Considers four objectives for cloud architectures
1. Up-gradation cost
2. Failure rate
3. Power consumption
4. Number of completed tasks
Does not consider any specialized many-objective optimization technique
Does not integrate any empirical performance characterization of clusters
Does not accumulate the impact of cooling energy consumption
33. Experimental Results: Improvement over PSO
and ACO in Real Testbed
• % of improvement over PSO in real testbed:
• % of improvement over ACO in real testbed:
Workload Time
Cooling
energy
Computation
energy
CPU usage
Memory
usage
18.9 GB 74 75 63 127 5
12.6 GB 36 35 25 0 -1
8.7 GB 21 21 20 10 -7
Workload Time
Cooling
energy
Computation
energy
CPU usage
Memory
usage
18.9 GB 49 49 58 159 0
12.6 GB -18 -20 10 27 3
8.7 GB 43 42 56 130 1
33
34. Impacts of Selecting A Threshold
34
Threshold value
Computation time
(mins)
Cooling energy
(KWh)
Computation
energy (KWh)
2 11.1 0.43 0.92
4 9.5 0.31 0.66
6 9.7 0.24 0.53
8 10.2 0.32 0.68
Simulation results with 67.7 GB data over 30 machines
Threshold value
Total no. of
solutions in the
pareto-front
No. of solutions
with < 6 machines
No. of solutions
with >= 6
machines
2 120 34 86
4 120 27 93
6 120 13 107
8 120 14 106
Too much small threshold can
result in worse results and
introduce unnecessary
solutions
36. Experimental Results: Improvement over PSO
in SimGrid
• % of improvement over PSO in SimGrid with 30 machines:
• % of improvement over PSO in SimGrid with 50 machines:
36
Workload Time Cooling energy
Computation
energy
67.7 GB 21 13 10
50.4 GB 36 11 10
28.3 GB 43 13 10
Workload Time Cooling energy
Computation
energy
67.7 GB 38 59 59
50.4 GB 33 36 38
28.3 GB 10 -6 0
37. Experimental Results: Improvement over ACO
in SimGrid
• % of improvement over ACO in SimGrid with 30 machines:
• % of improvement over ACO in SimGrid with 50 machines:
37
Workload Time Cooling energy
Computation
energy
67.7 GB 43 10 5
50.4 GB 17 8 8
28.3 GB 15 5 0
Workload Time Cooling energy
Computation
energy
67.7 GB 16 55 55
50.4 GB 41 79 79
28.3 GB 43 87 87
38. Experimental Setup for Validating Modified
NSGA-III Performs Better Than NSGA-III
38
Parameter Value
# of master machines 1
# of slave machines 29
Processor Intel Core 2 Duo
PC power Peak: 10 - 400 W, idle: 2.5-100 W,
power off: 5 W
Network B/W 10 to 100 kbps
Total # of files 86, 64, and 36
Size of each file 787 MB
Total data size 67.7 GB, 50.4 GB, and 28.3 GB
Maximum allowable temperature 20℃-23℃
Environment temperature 28℃
SimGrid version 3.12
40. Difference between Cluster and Grid?
40
Grid are inherently distributed by its nature over a LAN,
metropolitan or WAN. On the other hand, the computers
in the cluster are normally contained in a single location.
Another difference lies in the way resources are handled. In
case of Cluster, the whole system (all nodes) behave like a
single system view and resources are managed by
centralized resource manager. In case of Grid, every node
is autonomous i.e. it has its own resource manager and
behaves like an independent entity.
41. Difference between Grid and Cloud?
41
Grid can have only some high computing power.
Cluster can have some application level service.
FB, Google-> cloud service.
Cloud computing is a kind of grid computing;
it has evolved by addressing the QoS (quality of service)
and reliability problems.
42. Outline of Our Methodology
42
Cooling
energy in
SimGrid
Cooling energy in real
test-bed
Verify one another
Cooling energy integration
in SimGrid and test-bed
Modification
in NSGA-III
Selection
FilteringCrossover
Formulation and
modification in NSGA-III
to develop a solution
Compare
Modified
NSGA-III
ACOPSO
Performance
evaluation and
comparison
Problem formulation
43. Our Proposed Many-Objective Optimization
Model
43
Following the empirical analysis,
we formulate our objective
functions as follows:Computation
time
Energy
consumption
Cost
Resource
utilization (CPU
and memory
usage)
Workload
should be within
machine hard
disk size
No. of selected
machines
should be
greater than 0.
Cooling
temperature
Machine
configuration
Workload
No. of
machines
Temperature
difference
Machine
configuration
No. of
machines
Per machine
workload
Lowest
possible temp.
Max.
possible
temp.
44. Our Proposed Modification in NSGA-III
Algorithm
44
• With some probability we perform HUX
• Only perform when the decision variables are not
same
Half Uniform
Crossover (HUX)
• Cluster machines based on each machine property
• Make a group in which all machines are in best group
• Make a group in which all machines are in worst group
• Try to take from best group and not to take from worst
group
Greedy Clustering
Approach
• For each machine property try to take the best machine
instead of taking the worst machine based on probabilistic
possibility
Greedy Best and
Worst Swap
Modified crossover
45. Objectives and Possible Outcomes
45
Formulate a many-
objective optimization
problem for clusters,
and solve it
Integrating cooling
energy in both real
testbed and simulation
Test our approach in
real testbed and
simulation to compare
performance with
other approaches
Particle swarm
optimization (PSO)
Ant colony optimization
(ACO)
Modified
NSGA-III
46. Some Related Algorithms with Bio-Geography Based
Solutions
46
• Biogeography-based optimization (BBO)
• Ant colony optimization (ACO)
• Particle swarm optimization (PSO)
• Artificial Bee Colony Optimization
47. PSO Based Technique
47
1. Task assignment based on avg workload, the deviation of workload, and migration
cost.
2. PSO basic: Particle location is updated based on the current location, distance
between my current location and my best location, and distance between my current
location and global best location.
3. Limitations: Local optimal- can converge fast but local optima. Keep only
current optimal point.
Solution: Keep a global storage for keeping all the optimal values.
48. Ant Colony Based Technique
48
1. VM assignment based on resource wastage and power consumption.
2. ACO basic: Ants search for food. They put pheromone based on the quality of food
and amount. The other ants follow the path based on the concentration of pheromone.
49. BBO
49
• Define the mutation probability and the elitism parameter.
• Initialize the population
• Calculate the immigration rate and emigration rate for each island. Good solutions have
high emigration rates and low immigration rates. Bad solutions have low emigration rates
and high immigration rates.
• Probabilistically perform mutation for each island.
• Replace the worst islands in the population with the previous generation’s elite islands.
• If the termination criterion is met, terminate; otherwise, repeat
50. Stochastic Programming
50
Stochastic programming is a framework for modeling optimization problems that
Involve uncertainty.
Stochastic programming models are similar in style but take advantage of the fact
that probability distribution governing the data are known or can be estimated.
Stochastic programming studies the case in which some of the constraints or
parameters depend on random variables.
Stochastic linear programming is an extension of linear programming, where some
of the coefficients in the objective and the constraints are random variables
52. Modifications in NSGA-II to get NSGA-III
52
• Crowding distance idea has been replaced.
• NSGA-III introduces some reference points idea so that diversity is maintained.
53. Why we use NSGA-III as our baseline
algorithm?
Introduced in 2014
Specialized for solving many-objective optimization problems
Other earlier approaches mostly solve benchmark problems such as DTLZ. They
mostly ignore real-world problems
Popular NSGA-II only considers crowding distance, which is not a feasible
property in solving many-objective optimization problems
53
Popular MOEA/D claims to work for many-objectives, but only considered three
objectives only. Also, uses some default parameters
55. The Underneath of Cooling Energy Module Development for
Testbed Experiment: Circuit Diagram
55
1. Computation energy
2. Cooling energy
3. CPU usage
4. Memory usage
Energy
monitor
PT, CT
LM35
sensor
Psensor Energy monitoring module for measuring
computation power and cooling power
Temperature monitoring module
Circuit diagram
56. Experimental Results: SimGrid Energy Module: Simulation
Environment in Real Test-bed: Machine Configuration
Processor base
frequency
Memory # of machines
2.4 GHz 1 GB 1
2.4 GHz 2 GB 2
2.66 GHz 1 GB 2
2.66 GHz 2 GB 8
2.8 GHz 2 GB 17
56
58. How we get the equations? (cont.)
58
Using Matlab, we try to get the equations:
Computation time equations: (for 6GB data)
For order 1: 𝑦 = −0.562𝑥 + 17.2
For order 2: 𝑦 = 0.046𝑥2
− 2.098𝑥 + 26.8
For order 3: 𝑦 = −0.0038𝑥3
+ 0.2325𝑥2
− 4.73𝑥 + 36.4
Energy consumption equations: (for 6 GB data)
For order 1: 𝑦 = −0.0246𝑥 + .8950
For order 2: 𝑦 = 0.0023𝑥2
− 0.0996𝑥 + 1.3653
For order 3: 𝑦 = −0.0002𝑥3 + 0.0101𝑥2 − 0.2109𝑥 + 1.7722
Energy consumption equations: (for 6 GB data)
For order 1: 𝑦 = −1.4165𝑥 + 93.0156
For order 2: 𝑦 = 0.0094𝑥2 − 1.7271𝑥 + 94.9635
For order 3: 𝑦 = 0.0012𝑥3 − 0.0496𝑥2 − 0.8936𝑥 + 91.9160
59. Experimental Results: SimGrid Energy Module:
Cooling Energy and Computation Energy Ratio
• Cooling energy is around 83% of total energy in real
testbed
Workload
Cooling
energy
consumption
avg (KWh)
Computation
energy
consumption
avg (KWh)
Total energy
consumption
avg (KWh)
% of cooling
energy
consumption
% of
computation
energy
consumption
12.6 GB 0.367 0.088 0.455 81 19
9.44 GB 0.398 0.069 0.467 85 15
6.3 GB 0.242 0.052 0.294 82 18
3.14 GB 0.146 .030 0.176 83 17
59
60. Experimental Results: SimGrid Energy Module: Cooling
Energy and Computation Energy Ratio (Cont.)
• Cooling energy is around 65% of total energy in
SimGrid
Workload
Cooling
energy
consumption
avg (KWh)
Computation
energy
consumption
avg (KWh)
Total energy
consumption
avg (KWh)
% of cooling
energy
consumption
% of
computation
energy
consumption
12.6 GB 0.235 0.125 0.36 65 35
9.44 GB 0.176 0.094 0.27 65 35
6.3 GB 0.117 0.063 0.18 65 35
3.14 GB 0.058 .031 .089 65 35
60
61. Space vs Power Requirement?
61
We have 38*35 = 1330 sq ft room. Hence, we need 23,000 BTU/hr.
We have 3 ACs.
Each having capacity 1.5 ton = 18,000 BTU/hr.
Hence, 3 ACs will together give us 54,000 BTU.
63. Power Requirement for Different States
63
Power State Power
Active state 100W-120W
Idle state 40W
Power-off state 5W
Why we do not consider machine on off energy?
Answer: It is not a significant amount. We make the whole simulation for about 10-60
minutes. Machine on-off needs small amount of time. Hence, it should consume small
amount of energy. Moreover, we only consider the machine runtime energy measurement.
Hence, this on-off energy is not present in all the experiments. That is why the overall
comparison should not be affected.
65. Experimental Results: Improvement over PSO
and ACO in Real Testbed and SimGrid
• % of improvement over PSO comparison for 18.9 GB data in real
testbed and SimGrid:
• % of improvement over ACO comparison for 18.9 GB data in real
testbed and SimGrid:
Environment Time
Cooling
energy
Computation
energy
Real testbed 74 75 63
SimGrid 25 18 18
Environment Time
Cooling
energy
Computation
energy
Real Testbed 49 49 58
SimGrid 51 82 83
65
66. Experimental Results: Improvement over PSO
and ACO in Real Testbed and SimGrid
• % of improvement over PSO comparison for 31.5 GB data in real
testbed and SimGrid:
• % of improvement over ACO comparison for 31.5 GB data in real
testbed and SimGrid:
Environment Time
Cooling
energy
Computation
energy
Real testbed 43 43 44
SimGrid 28 56 58
Environment Time
Cooling
energy
Computation
energy
Real Testbed 28 9 28
SimGrid 36 82 82
66
68. Elbow Points from Our Year-Long Collected
Data to Guess The Threshold Value
68
Computation time with 6 GB data: Elbow point: 8 Computation time with 9 GB data: Elbow
point: 7
Total energy with 6 GB data: Elbow point: 10 Total energy with 9 GB data: Elbow point: 9
70. Selection Threshold Impacts over Different
Cluster Objectives
70
Threshold
value
Computation
time (mins)
Cooling energy
(KWh)
Computation
energy (KWh)
2 11.1 0.43 0.92
4 9.5 0.31 0.66
6 9.7 0.24 0.53
8 10.2 0.32 0.68
For a random cluster, we run our modified NSGA-III algorithm with different threshold
values with 67.7 GB data.
71. Selection Threshold Impacts over Solutions in
The Pareto-Front
71
Threshold
value
Total no. of
solutions in the
pareto-front
No. of
solutions with
< 6 machines
No. of
solutions with
>= 6 machines
2 120 34 86
4 120 27 93
6 120 13 107
8 120 14 106
For a random cluster, we run our modified NSGA-III algorithm with different threshold
values with 67.7 GB data.
Too much small
threshold can introduce
unnecessary solutions
72. Effects of Different Weights in Machine
Selection
72
Weights
Computation
time (mins)
Cooling
energy
(KWh)
Computation
energy
(KWh)
No. of
selected
machines
CPU
utilization
(%)
Memory
utilization
(%)
25-75 34.8 2.6 0.3 7 57.5 91.2
75-25 23.5 1.76 0.24 11 55.2 91.0
75-25 indicates more weights
to select highly configured
machines
25-75 indicates less weights
to select highly configured
machines
73. Experimental Results: Improvement over PSO
and ACO in Real Testbed and SimGrid
73
Graphs showing results for 18.9 GB and 31.5 GB workload
75. The Underneath of Cooling Energy Module Integration in
SimGrid: An Application of Heat Dynamics
We adopt the conventional heat formula:
𝑸 = 𝑪 𝒑 × 𝑾 × 𝑫𝑻
𝐶 𝑝 = Specific heat
𝑊 = Mass of airflow per min
𝐷𝑇 = Change in temperature
𝑸 = 𝑪 𝒑 × 𝑪𝑭𝑴 × 𝑫 × 𝑫𝑻
𝐶𝐹𝑀 = Cubic feet per minute
𝐷 = Density of the air
Thus, we can write:
Putting W = CFM × D, where CFM =
cooling air volume per minute, and
D = mass of air per volume, we get:
75
𝑪𝑭𝑴 =
𝑸
𝑪 𝒑 × 𝑫 × 𝑫𝑻
•Specific heat of a room remains constant
•At a certain place where the pressure is
constant, density will also remain constant
•There remain other heat losses
𝑪𝑭𝑴 =
𝟏. 𝟕𝟔 × 𝑸
𝑫𝑻
Using 1 CFM consumes 47.82W, we get:
𝑷𝒐𝒘𝒆𝒓 = 𝑪𝑭𝑴 × 𝟒𝟕. 𝟖𝟐
76. The Underneath of Energy Model and Utilization
Model Integration for Testbed Experiments
76
1. Computation energy
2. Cooling energy
3. CPU usage
4. Memory usage
Energy
monitor -
PT, CT
Psensor
Energy monitoring module for measuring
computation power and cooling power
Real testbed setup
77. The Underneath of Energy Model and Utilization
Model Integration for Testbed Experiment (Cont.)
77
PT
CT-1
CT-2
Energy Monitor 1
PT
CT-1
CT-2
Energy Monitor 2
Cluster voltage
Cluster
current
AC voltage
AC-2
current
Cluster
AC-1 AC-2 AC-3
AC-1
current
AC-3
current
78. Experimental Setup for Integration of Energy Module
in SimGrid and Real Testbed
Parameter
Value
SimGrid Testbed
# of master machine 1 1
# of slave machines 4, 9, 14, 19, 24, and 29
4, 9, 14, 19, 24, and 29
PC power
38, 400-44, 800
MFLOPS
2.4 GHz, 2.66 GHz, and
2.8 GHz
Machine power consumption
Peak: 100 - 120 W, idle:
40 W, power off: 5 W
N/A
Memory N/A 1-2 GB
Total # of files 4, 8, 12, and 16 4, 8, 12, and 16
Size of each file 787 MB 787 MB
Total data size
3.14. 6.3, 9.44, and 12.6
GB
3.14. 6.3, 9.44, and 12.6
GB
Maximum allowable temperature 22◦ Celsius 22◦ Celsius
Environment temperature 25◦ Celsius
25◦ Celsius
78
79. Experimental Results on Computation Energy
79
Computation energy comparison for
3.14GB data
Computation energy comparison for
6.3GB data
Computation energy comparison for
9.44GB data
Computation energy comparison for
12.6GB data
SimGrid computation energy and test-bed computation
energy follow a closely similar trend
80. Experimental Results on Cooling Energy
80
Cooling energy comparison for
3.14GB data
Cooling energy comparison for
6.3GB data
Cooling energy comparison for
9.44GB data
Cooling energy comparison for
12.6GB data
SimGrid cooling energy and test-bed cooling energy
follow a closely similar trend
81. Experimental Results on Total Energy
81
Total energy comparison for 3.14GB data Total energy comparison for 6.3GB data
Total energy comparison for 9.44GB data Total energy comparison for 12.6GB data
SimGrid outputs are close to real test-bed outputs
82. Exploiting a Synergy between Greedy Approach and NSGA
for Scheduling in Computing Clusters
Presented by
A S M Rizvi
This work was supported by the ICT Division, Government of the
People's Republic of Bangladesh.
A S M Rizvi – University of Southern California
Tarik Reza Toha – Bangladesh University of Engineering and Technology
Siddharthe Shankar Das – Bangladesh University of Engineering and Technology
Sriram Chellappan – University of South Florida
A. B. M. Alim Al Islam - Bangladesh University of Engineering and Technology
NSysS’2018
Notes de l'éditeur
Source: https://computing.llnl.gov/tutorials/parallel_comp/
Parallel computing is highly used for extensive computation. We can say about galaxy formation simulation or climate change prediction or weather forecast for future or even traffic simulation in busy roads. All these modeling, simulation and experimentation of complex real-world phenomena demands rigorous computing. Hence, we need parallel computing .
Source: https://computing.llnl.gov/tutorials/parallel_comp/
In parallel computing multiple processors work together to make faster the overall computation. In this slide, we can see that in the case of parallel computing we can solve three problems when we have parallel computing. However, for serial computing case, we can solve only one problem.
Clusters can have conflicting outcomes. If we increase the number of machines in a cluster, two conflicting results might appear. One is to decrease the computation time, and the other one is to increase the total cost. Decreasing computation time is desirable, however, increasing cost is not desirable. Hence, we have to make two optimization decisions. First we need to determine the number of machines that need to be operated. Secondly, we need to select which machines should be running. In this work, I try to give a solution for the cluster administrator to get answers of these two questions.
http://www.bmpcoe.org/library/books/ex-se/17b31.html
http://www.aitechnology.com/products/thermalinterface/
http://www.mtarr.co.uk/courses/ami4817_dti/u03/index.asp
These are the related works with our study. At first we can mention about the work of Tarplee. This work tries to optimize around four objectives. As we know, if any optimization problem has more than three objectives we know it as a many-objective optimization problem. However, in this work, authors use a multi-objective optimization technique to solve a many-objective optimization problem which might not give expected results.
Secondly, there are some works with multi-objective optimization in cloud infrastructure. However, cloud and cluster are different in architecture. Moreover, there are some works with optimization in cloud and cluster with PSO and ACO methods. These methods do not integrate any empirical performance characterization. Also, these techniques are not fully aware of cooling energy.
Weighted sum method
IJGDC means International Journal of Grid and Distributed Computing
These are the related works with our study. At first we can mention about the work of Tarplee. This work tries to optimize around four objectives. As we know, if any optimization problem has more than three objectives we know it as a many-objective optimization problem. However, in this work, authors use a multi-objective optimization technique to solve a many-objective optimization problem which might not give expected results.
Secondly, there are some works with multi-objective optimization in cloud infrastructure. However, cloud and cluster are different in architecture. Moreover, there are some works with optimization in cloud and cluster with PSO and ACO methods. These methods do not integrate any empirical performance characterization. Also, these techniques are not fully aware of cooling energy.
Weighted sum method
These are the related works with our study. At first we can mention about the work of Tarplee. This work tries to optimize around four objectives. As we know, if any optimization problem has more than three objectives we know it as a many-objective optimization problem. However, in this work, authors use a multi-objective optimization technique to solve a many-objective optimization problem which might not give expected results.
Secondly, there are some works with multi-objective optimization in cloud infrastructure. However, cloud and cluster are different in architecture. Moreover, there are some works with optimization in cloud and cluster with PSO and ACO methods. These methods do not integrate any empirical performance characterization. Also, these techniques are not fully aware of cooling energy.
Weighted sum method
Look at these two graphs.
From the first graph, we make the decision that if we increase the number of machines then the computation time will be decreased. From the second graph, we can say, if we increase the number of machines total energy is decreased. We make more understanding and relations with the cluster environment and cluster objectives and find these objective functions.
The first equation denotes computation time. It depends on machine configuration and workload.
Similarly, the second equation denotes the energy consumption……
Source: https://computing.llnl.gov/tutorials/parallel_comp/
In parallel computing multiple processors work together to make faster the overall computation. In this slide, we can see that in the case of parallel computing we can solve three problems when we have parallel computing. However, for serial computing case, we can solve only one problem.
Source: https://computing.llnl.gov/tutorials/parallel_comp/
In parallel computing multiple processors work together to make faster the overall computation. In this slide, we can see that in the case of parallel computing we can solve three problems when we have parallel computing. However, for serial computing case, we can solve only one problem.
These are the related works with our study. At first we can mention about the work of Tarplee. This work tries to optimize around four objectives. As we know, if any optimization problem has more than three objectives we know it as a many-objective optimization problem. However, in this work, authors use a multi-objective optimization technique to solve a many-objective optimization problem which might not give expected results.
Secondly, there are some works with multi-objective optimization in cloud infrastructure. However, cloud and cluster are different in architecture. Moreover, there are some works with optimization in cloud and cluster with PSO and ACO methods. These methods do not integrate any empirical performance characterization. Also, these techniques are not fully aware of cooling energy.
Weighted sum method
These are the related works with our study. At first we can mention about the work of Tarplee. This work tries to optimize around four objectives. As we know, if any optimization problem has more than three objectives we know it as a many-objective optimization problem. However, in this work, authors use a multi-objective optimization technique to solve a many-objective optimization problem which might not give expected results.
Secondly, there are some works with multi-objective optimization in cloud infrastructure. However, cloud and cluster are different in architecture. Moreover, there are some works with optimization in cloud and cluster with PSO and ACO methods. These methods do not integrate any empirical performance characterization. Also, these techniques are not fully aware of cooling energy.
Weighted sum method
For a random cluster, we run our modified NSGA-III algorithm with different threshold values with 67.7 GB data. (sIMULATION)
Look at these two graphs.
From the first graph, we make the decision that if we increase the number of machines then the computation time will be decreased. From the second graph, we can say, if we increase the number of machines total energy is decreased. We make more understanding and relations with the cluster environment and cluster objectives and find these objective functions.
The first equation denotes computation time. It depends on machine configuration and workload.
Similarly, the second equation denotes the energy consumption……
At first, we will try to formulate a many-objective optimization problem based on the empirical analysis of cluster environment. Then we will try to integrate cooling energy into real testbed and SimGrid. Then we will make some modification in the existing NSGA-III algorithm. We will test our approach in the real testbed and in SimGrid to compare with other approaches.
PT = Potential Transformer
CT = Current Transformer
PT = Potential Transformer
CT = Current Transformer
PT = Potential Transformer
CT = Current Transformer
PT = Potential Transformer
CT = Current Transformer
We use different weights to see how they manipulate the experimental results.In the first case, we put weights in such a way that smaller number of machines will be selected. Thus, we should have high computation time and energy. At the same time, lower number of machines should result in a decrease in the maintenance cost as well as increase in the CPU utilization. 25-75 weight indicates our algorithm selects high configured machines 25% time and 75-25 weight means our algorithm selects high configured machines 75% time on average. The following result is for 31.5 GB workload.
Ref.: A. Rizvi et al. Cooling energy integration in simgrid, NSysS, 2017
Ref: http://www.traditionaloven.com/tutorials/power/convert-atm-cfm-atmosphere-to-watts-w.html
Ref: http://www.sunon.com/uFiles/file/03_products/07-Technology/004.pdf
PT = Potential Transformer
CT = Current Transformer
PT = Potential Transformer
CT = Current Transformer
PT = Potential Transformer
CT = Current Transformer