Simulating Energy Aware Networks in Large Scale Distributed Systems - PowerPoint PPT Presentation

Simulating Energy Aware Networks in Large Scale Distributed Systems Betsegaw Lemma Amersho Supervised by: Prof. Martin Quinson, Dr. Anne-C´ ecile Orgerie Master Thesis Defence, June 30, 2017

Outline Recent Trends in Large-Scale Networks 1 Energy Proportionality 2 Common Research Methods 3 Network Simulation 4 Our Approach 5 Proposed Solution 6 Validation Experiments 7 Accuracy Validation Result Scalability Validation Result Conclusion 8

Recent Trends in Large-Scale Networks Networked Devices are Increasing 17.1 billion in 2016 to 27.1 billion in 2021 IP Traffic is Growing Global IP traffic will increase threefold over the next 5 years Cloud Infrastructure Dependency is Increasing By 2020, 92% of workloads will be processed by cloud data centers Data Center are Expanding In response to the added components and the increased service demand This Expansion Led to Energy Consumption Concern

Recent Trends in Large-Scale Networks Networked Devices are Increasing 17.1 billion in 2016 to 27.1 billion in 2021 IP Traffic is Growing Global IP traffic will increase threefold over the next 5 years Cloud Infrastructure Dependency is Increasing By 2020, 92% of workloads will be processed by cloud data centers Data Center are Expanding In response to the added components and the increased service demand This Expansion Led to Energy Consumption Concern

Recent Trends in Large-Scale Networks Networked Devices are Increasing 17.1 billion in 2016 to 27.1 billion in 2021 IP Traffic is Growing Global IP traffic will increase threefold over the next 5 years Cloud Infrastructure Dependency is Increasing By 2020, 92% of workloads will be processed by cloud data centers Data Center are Expanding In response to the added components and the increased service demand This Expansion Led to Energy Consumption Concern

Recent Trends in Large-Scale Networks Networked Devices are Increasing 17.1 billion in 2016 to 27.1 billion in 2021 IP Traffic is Growing Global IP traffic will increase threefold over the next 5 years Cloud Infrastructure Dependency is Increasing By 2020, 92% of workloads will be processed by cloud data centers Data Center are Expanding In response to the added components and the increased service demand This Expansion Led to Energy Consumption Concern

Energy Proportionality Ideal Energy Proportional Network Equipment Consume no power when idle (Static Power) Consume power in proportion to their work load (Dynamic Power) Energy Inefficiency of Current Network Equipment Current Network Equipment have narrow dynamic power range

Energy Proportionality Ideal Energy Proportional Network Equipment Consume no power when idle (Static Power) Consume power in proportion to their work load (Dynamic Power) Energy Inefficiency of Current Network Equipment Current Network Equipment have narrow dynamic power range

Common Research Methods Experimenting on Real Production Environment ( in vivo ) Can give real picture of the problem being studied Difficult to repeat experiments Might not be available for experimentation Experimenting on Experimental Test-Bed ( in vivo ) Offers full control over the experiment Limited in scalablility and for testing different scenarios Experimenting using Simulation Software ( in silico ) Gives full control, more flexible, less expensive and less time consuming Might fail to correctly model the real situation

Common Research Methods Experimenting on Real Production Environment ( in vivo ) Can give real picture of the problem being studied Difficult to repeat experiments Might not be available for experimentation Experimenting on Experimental Test-Bed ( in vivo ) Offers full control over the experiment Limited in scalablility and for testing different scenarios Experimenting using Simulation Software ( in silico ) Gives full control, more flexible, less expensive and less time consuming Might fail to correctly model the real situation

Common Research Methods Experimenting on Real Production Environment ( in vivo ) Can give real picture of the problem being studied Difficult to repeat experiments Might not be available for experimentation Experimenting on Experimental Test-Bed ( in vivo ) Offers full control over the experiment Limited in scalablility and for testing different scenarios Experimenting using Simulation Software ( in silico ) Gives full control, more flexible, less expensive and less time consuming Might fail to correctly model the real situation

Network Simulation Packet Level Simulators Modeling at a packet-level (capture low-level details) Are close to the real network phenomenon being modeld Fail to scale well when simulating large-scale networks Flow Level Simulators Abstract away low-level details and use analytical equations Suitable for simulating large-scale networks (Scalable) Loss of low-level details/accuracy The Goal of this Study To investigate the level of energy estimation accuracy that can be obtained from flow-level energy consumption models.

Network Simulation Packet Level Simulators Modeling at a packet-level (capture low-level details) Are close to the real network phenomenon being modeld Fail to scale well when simulating large-scale networks Flow Level Simulators Abstract away low-level details and use analytical equations Suitable for simulating large-scale networks (Scalable) Loss of low-level details/accuracy The Goal of this Study To investigate the level of energy estimation accuracy that can be obtained from flow-level energy consumption models.

Network Simulation Packet Level Simulators Modeling at a packet-level (capture low-level details) Are close to the real network phenomenon being modeld Fail to scale well when simulating large-scale networks Flow Level Simulators Abstract away low-level details and use analytical equations Suitable for simulating large-scale networks (Scalable) Loss of low-level details/accuracy The Goal of this Study To investigate the level of energy estimation accuracy that can be obtained from flow-level energy consumption models.

Our Approach Ⓐ Ⓒ Model Implementation Energy Consumption Model Literature Search Ⓔ Simulating n o Energy Consumption Model t i Ⓓ a d i l Validation a V l e Ⓑ d o M Comparing Scalability Simulating Ⓖ Energy Consumption Ⓕ

Proposed Flow Level Model Energy Consumption � T E ( T ) = P ( t ) dt 0 Power Consumption P total = P static + P dynamic Flow-Level Model Implemented in SimGrid � T E ( T ) = ( P static + P dynamic )( t ) dt where, 0 P dynamic = ( P busy − P idle ) ∗ u

Proposed Flow Level Model Energy Consumption � T E ( T ) = P ( t ) dt 0 Power Consumption P total = P static + P dynamic Flow-Level Model Implemented in SimGrid � T E ( T ) = ( P static + P dynamic )( t ) dt where, 0 P dynamic = ( P busy − P idle ) ∗ u

Proposed Flow Level Model Energy Consumption � T E ( T ) = P ( t ) dt 0 Power Consumption P total = P static + P dynamic Flow-Level Model Implemented in SimGrid � T E ( T ) = ( P static + P dynamic )( t ) dt where, 0 P dynamic = ( P busy − P idle ) ∗ u

Validation Experiments Accuracy Scenarios Data Size: [10,500] MB Traffic Flow: [1,4] Network Path Length: 1 and 3 Scalability Scenarios Simulation Time and Memory Usage Traffic Flow: 2 Data Size: [50,500] MB Network Path Length: 1, 2, 4, 6, 8, & 10

Validation Experiments Accuracy Scenarios Data Size: [10,500] MB Traffic Flow: [1,4] Network Path Length: 1 and 3 Scalability Scenarios Simulation Time and Memory Usage Traffic Flow: 2 Data Size: [50,500] MB Network Path Length: 1, 2, 4, 6, 8, & 10

Accuracy Validation Result ECOFEN vs SimGrid Comparison for all scenarios the accuracy estimation error is < 0.3% Flows 2, Path-Length 1 Flows 2, Path-Length 3

Scalability Validation Result Simulation Time and Peak Memory Usage Comparison SimGrid is 243 to 2723 times faster than ECOFEN SimGrid is 2 to 15 times more memory efficent than ECOFEN Peak Memory Usage Simulation Time

Conclusion Flow-Level energy consumption models can give energy estimation with very good accuracy without lossing their scalability.