with Different Service Levels and Quality of Service Andrei - - PowerPoint PPT Presentation

▶

Oct 18, 2022 130 likes •413 views

Scheduling for Cloud Computing with Different Service Levels and Quality of Service Andrei Tchernykh CICESE Research center Centro de Investigacin Cientfica y de Educacin Superior de Ensenada, Ensenada, Baja California, Mxico

SLIDE 1

Scheduling for Cloud Computing with Different Service Levels and Quality of Service

Aussois, France, March 31, 2014

Andrei Tchernykh

CICESE Research center

Centro de Investigación Científica y de Educación Superior de Ensenada, Ensenada, Baja California, México

chernykh@cicese.mx

http://www.cicese.mx/~chernykh

SLIDE 2

IPDPS 2012, IEEE 26th International Parallel and Distributed Processing Symposium

Uwe Schwiegelshohn University of Dortmund, Germany Andrei Tchernykh CICESE Research Center, Mexico

SLIDE 3

r1 r2 r3 time resource dj

Reserved Instances On-Demand Instances Spot-Batch Instances

ne-time payment for each instance you want to reserve

pay for compute capacity by the hour with no long-term commitments

If you have flexibility in when your applications can run

Amazon

SLIDE 4

Quality of Service

CICESE Parallel Computing Laboratory 4

rj dj

pj f1 . pj

f2 . pj

The provider guarantees to deliver the requested processing time within a certain time frame: slack or stretch factor fi

f1 =2: guarantees to deliver at least 50% of computing power

f2 =4: guarantees to deliver at least 25% of computing power

SLIDE 5

Quality of Service

CICESE Parallel Computing Laboratory 5

Deadline Service Level (slack factor and run time execution price) Execution time Profit

 Response time in relation to the requested processing time

SLIDE 6

Scheduling

CICESE Parallel Computing Laboratory 6

f=3

P1,2,3=4 P4=1

2 4 6 8 10 12

d2=d3=d4=12 d1=3 r1=r2=r3=r4=0

f=3 2 4 6 8 10 12

Optimal

SLIDE 7

Competitive Factor Obtained Income Optimal income

CICESE Parallel Computing Laboratory 7

Competitive Factor

𝑊∗ ≥ 𝑊 ∗ = 𝑣𝑛𝑏𝑦 ∙ 𝑛𝑗𝑜 𝑞𝑘

𝑜𝑠 𝑘=1

, 𝑒𝑛𝑏𝑦 ∙ 𝑛

SLIDE 8

CICESE Parallel Computing Laboratory 8

Scenarios

SSL-SM SSL-MM MSL-MM MSL-SM Number of different service levels: single service level or multiple service levels Number of machines: single machine model or multiple machine model

SLIDE 9

CICESE Parallel Computing Laboratory 9

Approach

Competitive analysis Worst case ratio between the obtained income and the optimal income over all instances: competitive factor Preemptive Earliest Deadline First scheduling First Fit allocation Greedy acceptance:

a job is accepted if possible.

Machine eligibility:

not all jobs are allowed to be executed to every machine.

Batch jobs (interactive jobs are not considered)

𝝇

SLIDE 10

Competitive Factor

CICESE Parallel Computing Laboratory 10

SSL-SM

𝝇 ≤ 𝟐 − (𝟐 −

𝒒𝒏𝒋𝒐 𝒒𝒏𝒃𝒚) 𝟐 𝒈 SSL-MM

𝝇 ≤ 𝒈 𝟐 + 𝒈(𝟐 − 𝒒𝒏𝒋𝒐 𝒒𝒏𝒃𝒚 )

Das Gupta and Palis, 2001

SLIDE 11

Competitive Factor

CICESE Parallel Computing Laboratory 11

𝝇 ≤ 𝒏𝒃𝒚{ 𝒒𝒏𝒋𝒐 𝒒𝒏𝒃𝒚 𝒈𝑱 − 𝟐 , 𝒈𝑱 − 𝟐 + 𝒒𝒏𝒋𝒐 𝒒𝒏𝒃𝒚 𝒈𝑱 − 𝟐 + 𝒗𝑱 𝒗𝑱𝑱

MSL-SM MSL-MM

𝝇 ≤ 𝒗𝑱𝑱 𝒗𝑱 (𝟐 − 𝟐 𝒈𝑱 )

SLIDE 12

Energy-Aware Online Scheduling: Ensuring Quality of Service for IaaS Clouds

Andrei Tchernykh CICESE Research Center, Mexico Luz Lozano Uwe Schwiegelshohn University of Dortmund, Germany Johnatan Pecero University of Luxembourg, Luxembourg Pascal Bouvry Sergio Nesmachnov Universidad de la República, Uruguay

SLIDE 13

Model extention

CICESE Parallel Computing Laboratory 13

𝑛 heterogeneous machines 𝑁𝑗 = (𝑡𝑗, 𝑓𝑔𝑔

𝑗)

– 𝑡𝑗 speed factor – 𝑓𝑔𝑔

𝑗

energy efficiency

𝑞𝑘

𝑗 = 𝑞𝑘/𝑡𝑗

processing time

𝑔

𝑗 = 𝑡𝑗

𝑡

𝑘

𝑡𝑗 is instruction execution speed 𝑡

𝑘 is execution speed guaranteed by SLA

SLIDE 14

Power consumption

CICESE Parallel Computing Laboratory 14

𝐹𝑝𝑞 = 𝑝𝑗 𝑢 ∙ 𝑄𝑗𝑒𝑚𝑓 + 𝑥𝑗 𝑢 𝑄𝑥𝑝𝑠𝑙

𝒏 𝑗=1 𝐷𝑛𝑏𝑦 𝑢=1

, 𝑝𝑗 𝑢 = 1, if on; 0 if off. 𝑥𝑗 𝑢 = 1 if work, 0 if idle.

SLIDE 15

Workload

CICESE Parallel Computing Laboratory 15

Site Proc s Power consumption W Energy efficien cy MFS/W Speed GFLOPS Log #Jobs #U ser 𝑸𝑱𝒆𝒎𝒇 𝑸𝑿𝒑𝒔𝒍𝒋𝒐𝒉 DAS2—University of Amsterdam 64 17.8 35.35 1.36 126 Gwa-t-1-anon_jobs- reduced.swf 1124772 333 DAS2—Delft University of Technology 64 17.8 35.35 1.36 126 DAS2—Utrecht University 64 17.8 35.35 1.36 126 DAS2—Leiden University 64 17.8 35.35 1.36 126 KTH—Swedish Royal Institute of Technology 100 17.8 26 1.75 18.6 DAS2—Vrije University Amsterdam 144 17.8 35.35 1.32 230 KTH-SP2-1996-2.swf 28489 204 HPC2N—High Perf. Comp. Center North, Sweden 240 58.9 66 0.89 481 HPC2N-2002-1.1-cln.swf 527371 256 CTC—Cornell Theory Center 430 17.8 26 1.64 88.4 CTC-SP2-1996-2.1- cln.swf 79302 679 LANL—Los Alamos National Lab 1024 24.7 31 1.45 65.4 LANL-CM5-1994-3.1- cln.swf 201387 211

speed, energy efficiency and power consumption of the machines and their workloads. range of the speed is [18.6, 481] efficiency [0.89, 1.75], power consumption [17.8, 58.9] for 𝑄𝑗𝑒𝑚𝑓, and [26, 66] for 𝑄𝑥𝑝𝑠𝑙.

SLIDE 16

Analysis

CICESE Parallel Computing Laboratory 16

Degradation in performance (relative error) of each strategy under each metric. This is done relative to the best performing strategy for the metric, as follows:

(𝛿 − 1) ∙ 100, with 𝛿 =

𝑡𝑢𝑠𝑏𝑢𝑓𝑕𝑧 𝑛𝑓𝑢𝑠𝑗𝑑 𝑤𝑏𝑚𝑣𝑓 𝑐𝑓𝑡𝑢 𝑔𝑝𝑣𝑜𝑒 𝑛𝑓𝑢𝑠𝑗𝑑 𝑤𝑏𝑚𝑣𝑓.

Performance profile. 𝜀(𝜐) is a non-decreasing, piecewise constant function that presents the probability that a ratio is within a 𝜐 factor of the best ratio. The function 𝜀(𝜐) is the cumulative distribution function. Strategies with large probability 𝜀(𝜐) for smaller 𝜐 will be preferred.

SLIDE 17

Strategies

CICESE Parallel Computing Laboratory 17 Type Strategy Level Description Knowledge-free

Rand 1 Allocates job 𝑘 to a machine randomly Ffit 1 Allocates job 𝑘 to the first machine available and capable to execute it. MLp 1 Allocates job 𝑘 to the machine with the least load at time 𝑠

𝑘: min 𝑗 =1…𝑛𝑏 𝑜𝑗 ,

Energy-aware

Max_eff 2 Allocates job 𝑘 to the machine with higher energy efficiency max

𝑗 =1…𝑛𝑏 𝑓𝑔𝑔 𝑗

Min_e 2 Allocates job j to the machine with minimum total power consumption at time 𝑠

𝑘 : min 𝑗 =1…𝑛𝑏

𝑄

𝑗 𝑝𝑞(𝑢) 𝑠𝑘 𝑢=1

MCT_eff 2 Allocates job 𝑘 to the machine with the earliest completion time on an energy efficient machine 𝑛𝑗𝑜

𝐷𝑛𝑏𝑦

𝑗

𝑓𝑔𝑔𝑗 , where 𝑑𝑛𝑏𝑦 𝑗

= max

𝑕𝑙=𝑗 𝑑𝑙 𝑗 and 𝑑𝑙 𝑗 being

the makespan and completion time of job 𝑙 in the machine 𝑗, respectively

Speed- aware

Max_seff 2 Allocates job j to the maximum energy efficient faster machine: max

𝑗 =1…𝑛𝑏 𝑡𝑗 ∗ 𝑓𝑔𝑔 𝑗

Max_s 2 Allocates job j to the fastest machine: max

𝑗 =1…𝑛𝑏 𝑡𝑗

SLIDE 18

Average income per machine SLA with 4 SLs

CICESE Parallel Computing Laboratory 18

0.E+00 5.E+07 1.E+08 2.E+08 2.E+08 3.E+08 3.E+08 4.E+08 4.E+08 1 2 4 8 16 32 64 128 income units machines FFit Max_eff Max_s Max_seff MCT_eff Min_e MLp Random

SLIDE 19

Total income, 4 SLs

CICESE Parallel Computing Laboratory 19

0.E+00 5.E+08 1.E+09 2.E+09 2.E+09 3.E+09 1 2 4 8 16 32 64 128 income units machines FFit Max_eff Max_s Max_seff MCT_eff Min_e MLp Random

SLIDE 20

Income degradation, 4 SLs

CICESE Parallel Computing Laboratory 20

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2 4 8 16 32 64 128 degradation machines Ffit Max_s Max_eff Max_seff MCT_eff Min_e MLp Random

SLIDE 21

Total power consumption, 4 SLs

CICESE Parallel Computing Laboratory 21

0.E+00 5.E+05 1.E+06 2.E+06 2.E+06 3.E+06 1 2 4 8 16 32 64 128 Wh machines FFit Max_eff Max_s Max_seff MCT_eff Min_e MLp Random

SLIDE 22

Degradation of the power consumption

CICESE Parallel Computing Laboratory 22

1 2 3 4 5 6 1 2 4 8 16 32 64 128 degradation machines Ffit Max_s Max_eff Max_seff MCT_eff Min_e MLp Random

SLIDE 23

Performance profile of the income, 8 strategies

CICESE Parallel Computing Laboratory 23

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 δ

𝜐

Random MLp Min_e MCT_eff Max_eff Max_seff Max_s Ffit

SLIDE 24

Performance profile of power consumption, 8 strategies

CICESE Parallel Computing Laboratory 24

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1 3 5 7 9 δ

𝜐

Random MLp Min_e MCT_eff Max_eff Max_seff Max_s Ffit

SLIDE 25

Solution space

CICESE Parallel Computing Laboratory 25

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 1 2 3 4 5 6 7 8 9 Income degradation Power consumption degradation

SLIDE 26

Pareto

CICESE Parallel Computing Laboratory 26

0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 Income degradation Power consumption degradation Random MLp Max_s Max_eff MCT_eff Min_e Max_seff FFit

SLIDE 27

Scheduling for Cloud Computing with Different Service Levels and Quality of Service

Aussois, France, March 31, 2014

Andrei Tchernykh

CICESE Research center

Centro de Investigación Científica y de Educación Superior de Ensenada, Ensenada, Baja California, México

chernykh@cicese.mx