[PPT] - IT Optimization Under Renewable Energy Constraint Gustavo PowerPoint Presentation

SLIDE 1

IT Optimization Under Renewable Energy Constraint

gustavo.rostirolla@irit.fr

Gustavo Rostirolla

Stephane Caux, Paul Renaud-Goud, Gustavo Rostirolla and Patricia Stolf. IT Optimization for Datacenters Under Renewable Power Constraint. Euro-Par 2018: Parallel Processing (Turin-Italy).

Advisors: Patricia Stolf and Stephane Caux

SLIDE 2

Agenda

Introduction
IT Optimization Module
Methodology
Results
Conclusion
Current and Future Work

Source: www.toulouse-tourisme.com

SLIDE 3

Introduction

Data centers are known as one of the big players when talking about energy consumption;
In 2006, were responsible for 61.4 billion kWh in the United States;
In 2010 about 1.3% of world's electricity;

3

Devices Networks Data Centers Manufacturing

20% 47% 15% 18% 29% 34% 21% 16%

Devices Networks Data Centers Manufacturing

2012 2017

Cook, Gary, et al. "Clicking Clean: Who is winning the race to build a Green Internet?." Greenpeace Inc., Washington, DC (2017).

SLIDE 4

Introduction

In the last years, the use of cloud computing has been the basis of data

centers, either in a public or private fashion.

4

Shehabi et al. 2016 - United States Data Center Energy Usage Report

~73 billion kWh in 2020

SLIDE 5

This migration to cloud computing increases the concern about power

utilization, especially when considering on site renewable energy sources and its oscillation over time;

Tasks submitted by users needs to be executed inside a time interval

(release time and due date);

But: When? Where? At which speed/frequency?

Introduction

5 Resource Time Energy

2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 Gigawatts 250 200 150 100 50

World Total

227 Gigawatts

+38 +38 +38 +40 +40 +50 +50 +29 +29 +30 +30 +17 +17 +17 +8 +6.5 +2.5 +1.4 +1.4 138 177 177 100 70 40 23 16 9 6.7 5.1 Annual Additions Capacity 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 Gigawatts 500 400 300 200 100

World Total

433 Gigawatts

318 +36 370 +52 +63 +63 238 +41 198 +39 159 +38 121 +27 94 +20 74 +15 59 +12 283 +45 +39 +38 +63 +27 +15 World Total World Total +45 +15 59 159 Annual additions Capacity Adib, Rana, et al. "Renewables 2016 global status report." Global Status Report Renewable Energy Policy Network for the 21st Century (REN21) (2016): 272.

SLIDE 6

Introduction - DataZERO

6

E l e c t r i c a l i n f r a s t r u c t u r e

H2 Cell

I T i n f r a s t r u c t u r e

Users Power

Decision Module Decision Module

Weather Forecast

New tasks

Negotiation

Electrical quality

0:00 6:00 12:00 18:00 Time of the day Power

Electrical quality

0:00 6:00 12:00 18:00 Time of the day Power 0:00 6:00 12:00 18:00 Time of the day Power

Common part Electrical power excess IT power excess Electrical proposal IT proposal Matching power plans

IT quality Electrical quality IT quality IT quality

Succesive Negotiation Rounds

www.datazero.org

SLIDE 7

SLIDE 8

✤ Set of Jobs J: Jj=[tj,i;pej,i;memj,i] ✤ tj,i=[trelase j,i;tduedate j,i;tdurationj,i] ✤ memj,i = Memory requested by task j,i ✤ pej,i = Number of processing elements requested by task j,i ✤ Set of Machines M: Mi=[npei;memi;Pmin i;Pmax i] ✤ [fi] set of frequencies available ✤ memi: Memory available in node ✤ Pmin i:Power when node is idle ✤ 𝛽i: Coefficient dependent on the machine ✤ npei: Number of processing elements ✤ Pmax i: g(Pmin i;fi,l;𝛽i) Power with processing element at 100% ✤ Discrete power curves Pavailable(t): power available at instant t

8

IT Optimization Module

w r d

Computing Node Processor 1 Core 1 Core 2 Core 3 Core 4 Processor 2 Core 1 Core 2 Core 3 Core 4 Memory

0 1 2 3 4 5 6 7 89...n

Time Power

SLIDE 9

1 2 3 4 5 6 7 8 9...n 1 2 3 4 5 6 7 8 9...n

9

✤ Output: ✤ Which task will run where,

when, at which frequency;

✤ Constraints {Power, CPU,

Memory}

✤ Translated as a set of

scheduling possibilities in the form of a power profile;

✤ Associated with metrics

(energy, due date violations…)

t4 t5 t3 t4

1 2 3 4 5 6 7 8 9...n

t4 t1 t5

Violations: 0 Violations: 1 Violations: 2 Energy: 7 Energy: 6 Energy: 6

t2 t5

Time (minutes) Power (W)

t1 t2 t1 t2

Power Metric Task

t3 t3

Time (minutes) Power (W) Time (minutes) Power (W)

IT Optimization Module

SLIDE 10

Meta-heuristic (Genetic

Algorithm):

Allows to produce a large

number of adapted solutions;

Makes it possible to approach

an optimum solution;

Slow execution time;
Difficulties in setting

parameters (crossover, mutation rate, population size, selection method).

10

Greedy Heuristic (Best Fit):
Fast scheduling decisions;
Easy implementation;
Tasks can be sorted by

arrival time, due date…;

Tasks are scheduled in a

local optimal, limited by the power curve received;

The combinations of choices

locally optimal do not always lead to an overall optimum.

IT Optimization Module

SLIDE 11

11

Generate Initial Population

Start

Calculate Fitness

f Individuals

Satisfy Stop Criterion

Selection Crossover Mutation

Greedy Algorithm to Set Tasks Start Time

Greedy Algorithm to adjust DVFS Calculate Fitness

f Children

Elitism for New Generation

End

Genetic Algorithm Workflow

No Yes

IT Optimization Module

SLIDE 12

12

Node 0 Node 0 Node 2 Node 3 T0 T1 T3 T2 Node 3 Node 3 Node 0 Node 2 T0 T1 T3 T2 Node 0 Node 0 Node 2 Node 3 T0 T1 T3 T2 Node 3 Node 3 Node 0 Node 2 T0 T1 T3 T2

Parents Offsprings

T0 T1 T3 T2 Node 3 Processor 1 Frequency: F0 Start: 1:00pm End: 2:00pm T0 T1 T3 T2 Node 3 Processor 2 Frequency: F1 Start: 1:00pm End: 2:00pm Node 0 Processor 1 Frequency: F0 Start: 3:00pm End: 5:00pm Node 2 Processor 1 Frequency: F3 Start: 1:00pm End: 2:00pm Node 0 Processor 1 Frequency: F0 Start: 1:00pm End: 2:00pm Node 0 Processor 2 Frequency: F1 Start: 1:00pm End: 2:00pm Node 2 Processor 1 Frequency: F0 Start: 1:00pm End: 2:00pm Node 3 Processor 1 Frequency: F3 Start: 2:00pm End: 3:00pm

Crossover/Mutation Greedy Time, Processor and Frequency

Genetic Algorithm:

IT Optimization Module

SLIDE 13

13

IT Optimization Module

Pr Pr Pr Pr T1 T3 T2 Node 0 Processor 1 Frequency: F1 Start: 1:00pm End: 5:00pm Node 0 Processor 2 Frequency: F0 Start: 1:00pm End: 2:30pm Node 0 Processor 2 Frequency: F3 Start: 2:30pm End: 5:00pm

(a)

DVFS:

T1 T2 Processor 1 Processor 2 T3

n
ff

T1 T2 Processor 1 Processor 2 T3

n
ff

(b) (c)

SLIDE 14

Two execution phases for Genetic Algorithm to improve execution time:
First phase provides an initial placement of tasks respecting a simplified

power curve;

Second phase uses the power prediction with all variations to improve this

initial placement, allowing the scheduling to take profit of power peaks.

14

1000 2000 3000 4000 5000 6000 7000 8000 100 200 300 400 500 600 Power (W) Time Step Refined Simplified

IT Optimization Module

SLIDE 15

Evaluation

SLIDE 16

Evaluation

✤ Computing Resources: ✤ Based in Paravance and Taurus (Grid5000) ✤ 30 Nodes x 2 Processors x 5 Frequencies

(1.2 to 2.4 Ghz);

✤ + Overhead of turning on/off a node. 16

Source: Villebonnet, V. (2016). Scheduling and Dynamic Provisioning for Energy Proportional Heterogeneous Infrastructures (Doctoral dissertation, Université de Lyon).

T. Mudge, “Power: A first-class architectural design constraint,” Computer, vol. 34, pp. 52– 58, 2001.

SLIDE 17

1000 2000 3000 4000 5000 6000 7000 8000 20000 40000 60000 80000 100000 120000 140000 160000 180000 Power (W) Time (seconds)

Winter Profile

Sun Wind Total

Evaluation

✤

2 days simulation with DCWorms:

✤

3 power profiles (real traces)

✤

3 workloads (Google based)

✤

234, 569 and 1029 tasks (known at beginning of execution)

17

1000 2000 3000 4000 5000 6000 7000 8000 20000 40000 60000 80000 100000 120000 140000 160000 180000 Power (W) Time (seconds)

Summer Profile

Sun Wind Total 1000 2000 3000 4000 5000 6000 7000 8000 20000 40000 60000 80000 100000 120000 140000 160000 180000 Power (W) Time (seconds)

Mixed Profile

Sun Wind Total

Profile II Profile I Profile III

SLIDE 18

Results

18

SLIDE 19

Results

19

Best Fit Due Date Best Fit Arrival Best Fit Size Genetic Algorithm Genetic Algorithm MO

15.81 50.19 78.11 15.86 50.32 77.24 15.69 50.38 77.03 14.84 45.35 68.14 14.91 45.41 67.87 14.85 45.63 67.49 14.62 45.07 67.65 14.69 45.21 67.30 14.55 45.15 67.03 18.34 53.26 81.03 18.09 53.92 81.24 17.47 53.78 81.39 14.79 46.53 73.18 14.94 46.12 73.48 15.40 47.04 73.15 0.00 30.00 60.00 90.00

234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks Profile I Profile II Profile III

Energy Consumption (kWh) 1 18 4 19 4 22 3 58 136 3 119 6 124 171 4 192 5 12 1 11 6 12 1 11 50 100 150 200

234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks Profile I Profile II Profile III

Due Date Violations

Violations Energy

SLIDE 20

Results

19

Best Fit Due Date Best Fit Arrival Best Fit Size Genetic Algorithm Genetic Algorithm MO

15.81 50.19 78.11 15.86 50.32 77.24 15.69 50.38 77.03 14.84 45.35 68.14 14.91 45.41 67.87 14.85 45.63 67.49 14.62 45.07 67.65 14.69 45.21 67.30 14.55 45.15 67.03 18.34 53.26 81.03 18.09 53.92 81.24 17.47 53.78 81.39 14.79 46.53 73.18 14.94 46.12 73.48 15.40 47.04 73.15 0.00 30.00 60.00 90.00

234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks Profile I Profile II Profile III

Energy Consumption (kWh) 1 18 4 19 4 22 3 58 136 3 119 6 124 171 4 192 5 12 1 11 6 12 1 11 50 100 150 200

234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks 234 Tasks 569 Tasks 1029 Tasks Profile I Profile II Profile III

Due Date Violations

Violations Energy

SLIDE 21

Results

20

Power Available Power Consumed

1000 2000 3000 4000 5000 6000 7000 8000 100 200 300 400 500 600 Power (W) Sample 1000 2000 3000 4000 5000 6000 7000 8000 100 200 300 400 500 600 Power (W) Sample

Best Fit Due Date Genetic Algorithm MO

✤

Examples of possible power consumption profiles of different algorithms. BF higher power consumption and number of violated tasks.

SLIDE 22

Results

20

Power Available Power Consumed

1000 2000 3000 4000 5000 6000 7000 8000 100 200 300 400 500 600 Power (W) Sample 1000 2000 3000 4000 5000 6000 7000 8000 100 200 300 400 500 600 Power (W) Sample

Best Fit Due Date Genetic Algorithm MO

✤

Examples of possible power consumption profiles of different algorithms. BF higher power consumption and number of violated tasks.

SLIDE 23

Interaction inside DataZERO

21 ITDM MQ Interface Message BUS Initial Profile Proposition Message Receive Profiles Response DCWorms MQ Interface DCWorms Scheduler

Power Time

Profile 1 Utility: 1000 Profile 2 Utility: 400 Profile 8 Utility: 500

. . .

SLIDE 24

Interaction inside DataZERO

22 DCWorms MQ Interface Output Description Datacenter Output Datacenter Activity Output Jobs Activity

Final Profile

DCWorms Scheduler

Power Time

Task Task Task Task Task Task Task Task Task Task Task Task Task Task Task Task T a s k T a s k Task Task Task Task Task Task T a s k Task Task Task Task

Power Time

SLIDE 25

Conclusion

Different algorithms that aims to minimize due date violations while

respecting power envelope and resource constraints;

Provide scheduling possibilities translated into power profiles with

associated metrics;

Reduction of both due date violations 192 (Greedy) to 11 (GA) and

energy 81.39 kWh to 73.15 kWh.

Integration between power production and IT load;

23

SLIDE 26

Current Work

Scheduling of mixed workload batch and services;

24

Release Duedate Duration

Service Batch

Start Time Unknown End Time

Service Base Hardware: 4 Cores 2.3 GHz 10GB Memory CPU (%) Memory (MB) I/O (kbps)

Phases based tasks generator.

P1 P2d P3 P4 P5d

Normal Service Phases (Time) Degraded Service Resources (%)

SLIDE 27

Current Work

25

SLIDE 28

Current Work

26

Price range per amount of data

used

Availability Ratio Credit 99.95% to 99.00% 10% 99.00% to 95.00% 25% 95.00% to 90.00% 35% <90.00% 50%

Data Movement Price Range Cost per GB <1 GB $0.00 0.001-10 TB $0.09% 10-50 TB $0.085 50-150 TB $0.07 150> TB $0.05 Violation Ratio Credit 0 to 5% 5% + Flex j/100 5% to 10% 10% + Flex j/100 10% to 20% 20% + Flex j/100 > 30% 30% + Flex j/100

Compensation for service

degradation

Compensation for batch violation

and user flexibility (in case of violation)

SLIDE 29

Current Work

27

0.00 0.25 0.50 0.75 1.00 12:00 Time Task 1 2 3 4 5 6 7 8 9 10 11 12 18:00 Total CPU Usage 00:00 06:00 0.00 0.25 0.50 0.75 1.00 Task 1 2 3 4 5 6 7 8 9 10 11 12 Total CPU Usage 00:00

Power Constraint

12:00 06:00 Time 18:00

SLIDE 30

Future Works

Evaluate phases based scheduling algorithms
Weather events (online)
Choose algorithm based on input (time, workload…)
Integration with other modules.

28

SLIDE 31

IT Optimization Under Renewable Energy Constraint

gustavo.rostirolla@irit.fr

Gustavo Rostirolla

Stephane Caux, Paul Renaud-Goud, Gustavo Rostirolla and Patricia Stolf. IT Optimization for Datacenters Under Renewable Power Constraint. Euro-Par 2018: Parallel Processing (Turin-Italy).