Evaluation on performance and energy eciency of distributed computing systems

Moreover,energy efficiency characteristic based scheduling can conserve the system energy and highperformance priority based policy yields the best performance.. Second, new real-time me

Efficiency of Distributed Computing

Systems

Ph.D Dissertation

byTran Thi Xuan (MSc)

Supervised byProf Do Van Tien (DSc)Department of Networked Systems and ServicesBudapest University of Technology and Economics

Hungary, 2020

The increasing usage of distributed computing systems to serve the growing demand forscientific computation and big data processing comes with the drastic growth of energyconsumption in computing clusters Therefore, optimizing the energy consumption ofcomputational clusters has become more crucial than ever The dissertation summarizes

a study on the resource allocation problem in distributed systems, motivated by a need

of taking into account different resource characteristics and dynamic power management(DPM) techniques

First, a generalized model of computational clusters built from heterogeneous types

of COTS servers has been introduced to study the resource-aware scheduling A set ofscheduling heuristics that consider servers’ performance and power consumption charac-teristics and the organization of waiting buffers have been investigated We show that thebuffering schemes play an important role in ensuring the quality of service parameters interms of the waiting time and the response time experienced by arriving jobs Moreover,energy efficiency characteristic based scheduling can conserve the system energy and highperformance priority based policy yields the best performance

Second, new real-time measurement based scheduling algorithms to achieve a trade-offbetween energy efficiency and the performance capability of computational clusters havebeen proposed in the thesis Numerical results show that the proposed algorithms attain

a balance between the job execution time and energy efficiency

Third, the impact of dynamic power management (DPM) in computing systems builtfrom multicore processors has been investigated Numerical results point out that DPM

in the core level of processors can play a role in saving energy consumption A aware scheduling solution has been proposed to achieve energy-efficient processing ofparallel tasks in multicore systems Obtained results indicate that the proposal reducesenergy consumption significantly in comparison to random allocation

resource-2

Last, the energy inefficiency in an ordinary big data scheduler-Hadoop YARN hasbeen investigated Since the resource allocation policy in the Hadoop YARN cluster isdata-aware (i.e the allocation strongly depends on the locations of data splits in HadoopDistributed File System-HDFS), a new data placement scheme for HDFS was proposed toachieve energy efficiency when MapReduce tasks are processed by the cluster Compared

to the existing HDFS data layout scheme, the proposal yields above 50% reduction inenergy consumption at a small expense of ≈6% increase in job execution time

I, the undersigned Tran Thi Xuan, hereby state that I have written this doctoral sertation myself, and I have used only the sources given in it I have clearly marked allthe parts taken from other sources either word for word or reworded but with the samecontents, indicating their sources.

dis-The reviews of the dissertation and the report of the thesis discussion are available at theDean’s Office of the Electrical Engineering and Informatics Faculty, Budapest University

of Technology and Economics

Budapest, February 17, 2020

Tran Thi Xuan

I would like to thank all people who have provided invaluable assistance during my studytowards the Ph.D degree.

I would like to express my sincere gratitude to Prof Dr Do Van Tien for his intensivesupervision Prof Dr Do Van Tien has guided me on the direction of my research atpreliminary time Without his continuous supervision and straight criticisms, I could notaccomplish this study and achieve PhD degree

I deeply thank Dr Do Hoai Nam, a senior researcher in Analysis, Design and velopment of ICT systems laboratory at our department, for his work cooperation andenthusiastic support through my research All members of the Analysis, Design and De-velopment of ICT systems laboratory, other PhD students, and the university staffs areacknowledged

De-Finally, I dedicate my hearty thankfulness to my husband and son Le Linh Bang and

Le Minh Anh for their love and encouragement I am also grateful to all family membersand friends who have supported me throughout

6

Abstract 4

2.1 Introduction 20

2.2 A generalized cluster model and Scheduling algorithms 21

2.2.1 Ranking of servers 22

2.2.2 Scheduling algorithms 23

2.2.3 Performance measures and energy metrics 27

2.3 Simulation Inputs and Numerical Results 29

2.3.1 Input parameters 29

2.3.2 Numerical results 31

2.4 Conclusion 39

8

3 New algorithms for balancing energy consumption and performance in

3.1 Introduction 41

3.2 System description and proposed scheduling algorithms 42

3.2.1 Scheduling algorithms 42

3.3 Numerical Results 45

3.3.1 The parameters of a computational cluster 46

3.3.2 Job balance 47

3.3.3 System metrics 48

3.3.4 Impacts of DVFS 51

3.3.5 Evaluations with workload traces as input data 52

3.4 Conclusion 55

4 Impact of Dynamic power management techniques in computing sys-tems of multicore processors 56 4.1 Introduction 57

4.2 Dynamic Power Management practices 58

4.3 System descriptions and operation scenarios 59

4.3.1 Job assignment scenarios 61

4.3.2 Performance and energy metrics 63

4.4 Evaluation on the impact of DPM 65

4.4.1 Simulation inputs 65

4.4.2 Analysis of obtained results 67

4.5 A proposal of Resource-aware scheduling algorithm 72

4.5.1 The proposed policy 73

4.5.2 Numerical results 75

4.6 Conclusion 79

5 A New Data Layout Scheme for Energy-Efficient MapReduce Processing Tasks 80 5.1 Introduction 81

5.2 Related Work 81

5.3 The operation of HDFS and YARN in a computing cluster 83

5.3.1 HDFS 83

5.3.2 Yet Another Resource Negotiator –YARN 83

5.3.3 Processing Hadoop MapReduce applications 85

5.3.4 The default HDFS data layout 86

5.3.5 A locality relaxation algorithm for resource allocation in RM 87

5.4 A New Data Layout Algorithm 89

5.4.1 Subsets of servers 89

5.4.2 A proposed algorithm 90

5.5 Numerical Results 92

5.5.1 Parameters for a case study 93

5.5.2 Numerical results 98

5.6 Conclusion 103

2.1 Separate Queue 25

2.2 Class Queue 25

2.3 Common Queue 26

2.4 Mean service time vs system load 31

2.5 Mean response time vs system load 32

2.6 Mean waiting time vs system load 32

2.7 CDF of response time at U = 50% 33

2.8 CDF of response time at U = 90% 33

2.9 AEno−switch vs system load 34

2.10 AEswitch−of f vs system load 34

2.11 Mean energy consumption vs system load 35

2.12 Average service time vs load 36

2.13 Average Energy Consumption (with switch-off ) vs load 36

2.14 Response time vs system load 37

2.15 Mean energy consumption vs system load 38

2.16 Response time vs λ 38

11

2.17 CDF of Response time 39

2.18 Mean energy consumption vs λ 39

3.1 Job distribution vs system load (EE policy) 47

3.2 Job distribution vs system load (HP policy) 47

3.3 Job distribution vs load (Real-Load algorithm) 47

3.4 Job distribution vs load (AVR_ST) 47

3.5 Mean service time vs system load 49

3.6 Mean waiting time vs system load 49

3.7 Mean response time vs system load 49

3.8 Energy consumption (switch-off) vs system load 50

3.9 Energy consumption (no switch-off) vs system load 50

3.10 System metrics with applied DVFS 51

3.11 Energy saving ratio vs λ 52

3.12 Degradation of response time vs λ 52

3.13 Q-Q plot for INCC workload 52

3.14 Q-Q plot for UPR workload 53

3.15 Energy consumption (switch-off) with two input data traces 53

3.16 Mean waiting time with two input data traces 54

3.17 Mean response time with two input data traces 54

4.1 A heterogeneous cluster model of multicore servers 60

4.2 Mean waiting time per job vs scheduling scenarios 68

4.3 Mean response time per job vs scheduling scenarios 69

4.4 System throughput vs scheduling scenarios (case of ICCN workload) 69

4.5 AIE vs scheduling scenarios 70

4.6 AAE vs scheduling scenarios 71

4.7 AOE vs scheduling scenarios 72

4.8 A gang job example: job j is split into T aj = 3 sibling tasks which must be processed simultaneously 73

4.9 Average operating energy consumption (KW.s/job) 77

4.10 Average waiting time (s/job) 78

4.11 Average response time (s/job) 78

4.12 Measured loads vs theoretical loads 78

5.1 A configuration of HDFS and YARN components in a computational cluster 84 5.2 An example of MapReduce WordCount execution on a Hadoop computing cluster 86

5.3 An example of data placement (with replication factor RF=3) in the ranked cluster 91

5.4 A completion time model for MapReduce workload 96

5.5 Average number of busy servers - with a Uniform dist 99

5.6 Average idle energy consumption (kJ/job) - with a Uniform dist 99

5.7 Average energy consumption (kJ/job) - with a Uniform dist 99

5.8 Average operating energy consumption (kJ/job) - with a Uniform dist 100

5.9 Average operating energy consumption (kJ/job) - with a Normal dist 100

5.10 Mean computation time - with a Uniform dist 101

5.11 Mean response time - with a Uniform dist 101

5.12 Mean response time - with a Normal dist 1025.13 ECDF of the execution time of jobs - with a Normal dist 102

2.1 System of notations 29

2.2 Server specifications 30

2.3 Ranking values 30

3.1 Statistics of workload during operations 45

3.2 Server specifications 46

3.3 Ranking function 46

3.4 Server specifications at 70% workload 51

4.1 Notations 64

4.2 Server parameters 65

4.3 Ranks of servers 66

4.4 Statistics of traces 66

4.5 Performance capacity and service rate of a CPU core 67

4.6 Server parameters 76

4.7 Capacity per core 76

5.1 System of Notations 92

15

5.2 Machine hardware configuration 93

5.3 Machine parameters 93

5.4 Capacity per core 95

5.5 Performance measures 98

Distributed computing plays a key role in solving computational scientific problems as well

as processing large volumes (terabytes) of data [1, 2, 3, 4] With the rapidly increasingscale of such systems, minimizing operation energy consumption has become more crucialthan ever Various approaches have been adopted to reduce energy consumption in ITindustry

Dynamic power management (DPM) practices are switching off idle servers and namic Voltage and Frequency Scaling (DVFS) [14, 19, 37] Energy optimization usingswitching off technique has received considerable interest since the energy consumed inidle state contributes a non-negligible fraction in the total energy cost [18, 44, 45, 47].DVFS reduces power consumption of active processors by scaling up/down the operat-ing voltage and frequency of CPU This technique, however, comes with degradation insystem performance and needs careful consideration before applied [17, 49]

Dy-Numerous researches focused on energy-efficient scheduling policies with the aim atgaining the efficiency of DPM practices in computing systems [48, 49, 50, 18, 33, 39].With the continuous change of hardware design as well as the system infrastructure,the application of DPM remains an open research question In addition, several studiespointed out that resource heterogeneity (processors have different power characteristicsand performance capacities) can be taken into scheduling decision to achieve energy con-servation [91, 18, 41, 51, 84, 85, 86] Those studies, however, lack of full consideration forthe heterogeneity of resource characteristics and a solution to balance the system perfor-mance and energy efficiency Furthermore, the resource heterogeneity and DPM practiceshave not been investigated adequately in big data processing systems

17

The dissertation summarizes a study on the resource allocation problem in distributedsystems, motivated by a need of taking into account different resource characteristics andDPM techniques.

First, a generalized model of computational clusters built from heterogeneous types

of COTS servers has been introduced to study the resource-aware scheduling A set ofscheduling heuristics that consider servers’ performance and power consumption charac-teristics and the organization of waiting buffers have been investigated The results andcontributions are presented in Chapter 2

Second, new real-time measurement based scheduling algorithms to achieve a off between energy efficiency and the performance capability of computational clustershave been proposed in the thesis Numerical results show that the proposed algorithmsattain a balance between the job execution time and energy efficiency The results andcontributions are presented in Chapter 3

trade-Third, the impact of dynamic power management (DPM) in computing systems builtfrom multicore processors has been investigated Numerical results pointed out that DPM

in cores of processors can play a role in saving energy consumption A resource-awarescheduling solution has been proposed to achieve energy-efficient processing of paralleltasks in multicore systems The contributions are presented in Chapter 4

Last, the energy inefficiency in an ordinary big data scheduler-Hadoop YARN has beeninvestigated Since the YARN scheduler leans on the data locality optimization to performthe resource allocation task, a new data placement scheme for Hadoop Distributed FileSystem (HDFS) was proposed The contributions are presented in Chapter 5

A generalized model of heterogeneous computing clusters for investigation of scheduling schemes

19

2.1 Introduction

The advance of high-speed networking and powerful computers together with the rapiddecline of hardware costs led to the widespread application of distributed systems tooffer services [22, 82] In such computational systems, job scheduling is of multi-criteria

in nature and is a vital task in state-of-the-art resource allocation studies The rapidincrease in the size and complexity of computational clusters has a significant impact onthe energy consumption, which is to be considered in the operation of systems

In the literature, job allocation algorithms have been proposed to schedule arrivingjobs in computational clusters (see [74, 73, 87] and references therein) In addition, somealgorithms are designed with regard to the knowledge about characteristics of jobs Thesemay belong to either clairvoyant [90, 89] or non-clairvoyant algorithms [88] Nowadays,optimizing energy consumption has become as crucial as improving performance [18, 14]

To reduce the energy consumption of modern Commercial Off-The-Shelf (COTS) servers,several power management (PM) practices can be applied: switch-off technique puts server

in very low-power sleep states by shutting down the idle components (e.g., disk, CPU),while Dynamic Voltage and Frequency Scaling (DVFS) lowers the operating voltage/fre-quency of CPU Furthermore, the energy consumption of computational clusters can bereduced if the operator systematically exploits the resource heterogeneity where manytypes of processors with different power characteristics and performance capacities areconsisted This was the main motivation of the investigations performed by Zikos andKaratza [91], where three policies applicable for cluster-level scheduling were compared:SQEE (Shortest Queue based policy with Energy Efficiency priority), SQHP (ShortestQueue based policy with High Performance priority), and PBP-SQ (Performance-BasedProbabilistic - Shortest Queue) Their simulation results indicated that SQEE is thebest from the aspect of energy consumption, SQHP outperforms the other two schemeswith the price of higher energy consumption, and PBP-SQ is the worst among the threeschemes In addition, Terzopoulos and Karatza [74], Gkoutioudi and Karatza [25] alsoinvestigated the scheduling in real-time grid systems

In this work, we follow the same approach applied in the study by Zikos and Karatza[91], where compute-intensive jobs with unknown service times are to be scheduled andexecuted in a cluster with heterogeneous servers Compared to previous works [91, 74, 25],this work provides a generalized model of heterogeneous types of COTS servers A method

to rank physical servers based on either high performance priority or energy efficiency ority, and measures to characterize the performance of computational clusters are defined

pri-We also investigate three schemes (Separate Queue, Class Queue and Common Queue)

for buffering jobs in a computational cluster These proposals allow a systematic way toinvestigate the performance of computational clusters.

Simulation results show that the buffering schemes do not have impact on performancemeasures related to the energy consumption of the investigated cluster, but are significantfactors regarding the waiting time and the response time experienced by arriving jobs Inaddition, we also investigate the energy consumption when a specific server is switched off

if there is no job allocated to the specific server and when a specific server applies DVFS

if a job is allocated to the server Results show that DVFS should be carefully applied toreduce the energy consumption of computational clusters

The rest of the chapter is organized as follows In Section 2.2, a generalized model,methods to rank physical servers, and scheduling algorithms are defined Simulationresults are presented in Section 2.3 Finally, Section 2.4 concludes the study

We consider a computational cluster, which serves compute-intensive jobs Following [91],

we assume that jobs:

• can be executed on any server,

• are attended to by the First Come First Served (FCFS) service policy,

• are non-preemptible, which means they cannot be suspended until completion,

• have service times unknown to the local scheduler

Jobs are to be executed by physical servers according to a specific scheduling algorithm

In what follows, we describe scheduling algorithms that allocate arriving jobs to a putational cluster To make the presentation comprehensible, the classification of servers

com-is provided in Subsection 2.2.1 Then, the description of scheduling algorithms com-is given

in Subsection 2.2.2

2.2.1 Ranking of servers

In a computational cluster, each physical server belongs to a specific server type Let S

interval divided by the number of seconds defined for this interval, showing thethroughput - workload operations per second - for this period at 100% target load);

We assume that when a server is busy, it functions at the full load with the power

We introduce two ranking functions based on servers’ parameters as follows:

s, which denotes the energy efficiency characteristics of servers in class s

energy efficiency (EE) priority ranking),

s ∈ S, respectively;

• number 1 is assigned to server type arg max

• number |S|, assigned to server type arg min

Performance Evaluation Corporation (SPEC)

For ranking Sp , if there are two server types with the same ssj_ops value, the servertype of a higher average active power gets a higher index We organize physical servers into

K classes based on their type Physical servers of the same type form a class, hence K =

servers are indexed by pair (i, j), (i = 1, , K; j = 1, , M(i)) Note that server (1,j), j = 1, , M(1), has the highest priority and server (K, j), j = 1, , M(K), has thelowest priority

The task of scheduling algorithms is to allocate arriving jobs to physical servers ing algorithms can take into account several factors such as the performance, the powerconsumption, the number of waiting jobs Furthermore, the organization of waiting spacefor jobs that are not immediately served upon their arrival is an important question

Schedul-In [91], the authors assumed that an arriving job will wait in a specific physical serverafter the scheduling decision, which is quite straightforward from the aspect of implemen-tation We call this queuing solution as a Separate Queue Scheme which is illustrated inFig 2.1 We also investigate two further schemes for buffering jobs

• Separate Queue Scheme: An arriving job will wait in a specific physical server afterthe scheduling decision presented in Algorithm 1.s one observes, the schedulingalgorithm chooses the server with the shortest queue If there are more idle servers

or servers of the same queue length, a server is selected based on the priority of itsserver type

• Class Queue Scheme: A buffer is associated to each class of servers (Figure 2.2).Jobs scheduled to a specific class wait in the buffer when all servers of the classare busy When a job departs from any physical server, the first waiting job in thebuffer of the same class is routed to that server Algorithm 2 routes an arriving jobbased on criteria: idle servers, the priority and the shortest queue length of classes

• Common Queue Scheme: there is one buffer for storing jobs (illustrated in ure 2.3) On a job arrival, if the LS finds all servers busy, the job will be stored inthe common queue and will wait for an idle server A job is immediately served ifAlgorithm 3 finds an idle server upon its arrival

Fig-Algorithm 1 The scheduling algorithm for Separate Queue

SCHEDULE: ROUTE job to queue of server best_server

Algorithm 2 The routing algorithm for Class Queue

Figure 2.1: Separate Queue

Figure 2.3: Common Queue

Algorithm 3 The routing algorithm for Common Queue

if found f ree_server then

ROUTE job to f ree_server

else

ROUTE job to Common Queue

end if

It is worth emphasizing that the practical implementation of the Separate QueueScheme is easiest That is, waiting jobs can be placed inside each physical server Forexample, jobs and parameters can be allocated in the local disk of each physical server.

To implement the Separate Queue Scheme, the Class Queue Scheme and the CommonQueue Scheme we propose a practical method as follows

• A file server is operated in a cluster that accessed by Server Message Block/theCommon Internet File System (SMB/CIFS) protocol [30] or Network File System(NFS) protocol [15]

• Each queue is allocated a separate spool area/directory in the file server The spoolareas are accessible by the Local Scheduler (LS) and respective servers

• The LS maintains the communication with the physical servers in the cluster withthe help of a Grid Computing Framework that allows loading and executing tasks.The LS also has the knowledge on the information of the occupancy of each queueand the states of physical servers in the cluster Based on the knowledge and thecommunication mechanism and the common spool areas, waiting jobs can be easilyallocated to servers depending the applied buffering approaches The allocation can

be performed quickly in the fast local network of the cluster, which minimally affectsthe performance of the cluster

needed to process job l The mean waiting time W T (n) and mean service time ST (n) of

n completed jobs are calculated as follows:

Response time rlof job l is the time period between the arrival instant and the departure

, the average energy consumption per job when idle servers are not switched off (butfunctioning in the idle state with a lower power) until n jobs finishes is

Table 2.1: System of notations

AEswitch−of f average energy consumption per job with switching-off

AEno−switch average energy consumption per job with no switching-off

We have implemented the buffering sheme based system models with appropriate ing algorithms described in Section 2.2 using SimPack simulation toolkit [21] In whatfollows, if it is not stated otherwise, our simulations are performed with the confidencelevel of 99% The accuracy (i.e the ratio of the half-width of the confidence interval and

the accuracy of waiting time is less than 0.001

We assume that the considered computational cluster is built from three types of mercial Off-The-Shelf (COTS) servers These types are listed in Table 2.2

Com-The computing capability (performance) and the power consumption of three server types

Pid,s is the power consumption when the servers are in the idle state The computed

rankings based on the rule in Eqs (2.1) and (2.2) are reported in Table 2.3

Job arrivals follow an exponential distribution with parameter λ The execution times of

jobs are also exponentially distributed Each job requires a computing capacity equivalent

to 6419253 (ssj_ops) in average, which means the average execution time of jobs is one

the service rates, if jobs are scheduled to a server of type Intel Xeon E5-2670,Intel Xeon

required by jobs and Table 2.2

We investigate a computational cluster with three server classes; each has eight

computational cluster is 8 ∗ (6419253 + 5286503 + 2790966) = 14496722 The

inves-tigation is carried out with load equal to 50%, 60%, 70%, 80% and 90%, where U =

λ/(8 ∗ (6419253 + 5286503 + 2790966))/6419253 = λ/(8 ∗ (1 + 0.82 + 0.43)) Therefore,

the mean inter-arrival time of jobs in our simulation runs takes the values of 0.1111 (s),

0.0926 (s), 0.0794 (s), 0.0694 (s) and 0.0617 (s), respectively

2.3.2 Numerical results

In this section, we present results concerning the three buffering schemes with two ing policies based on the energy efficiency (EE) priority and the high performance (HP)priority

Figure 2.4 illustrates the mean service time against the system load for all system modelsand policies The results show that tendencies observed in the case of the EE and the HPpriority are opposing: the operation under the EE policy yields a shorter service time asthe system load increases, while the HP policy results in a longer service time Howeverthey all converge to the same service time rounded value of 1.35 (s) at high load

This can be explained by the dominant utilization of high performing servers in case ofthe high system load under the EE policy, and generally under the HP policy A betterperforming server requires a shorter amount of time to process the exact job as a lowerperforming, but less energy demanding server The anticipation that the mean servicetimes are not impacted by queueing schemes is confirmed by Figure 2.4

The mean response time is depicted in Figure 2.5 It can be seen that the two proposed

(a) EE policy applied

1 1.2 1.4 1.6 1.8

50% 60% 70% 80% 90%

U

Separate-Queue Class-Queue Common-Queue

(b) HP policy applied

models outperform the traditional model under both policies and at every system load

In terms of response time, the Common Queue performs the best, especially under highsystem utilization The difference in performance increase directly with system load:while no significant difference can be discerned at 50%, a noticeable gap appears at 90%

50% 60% 70% 80% 90%

U

Separate-Queue Class-Queue Common-Queue

50% 60% 70% 80% 90%

U

Separate-Queue Class-Queue Common-Queue

(b) HP policy applied

The differences in mean response times can be resulted in from the expected waitingtime that job spends in system, shown in Figure 2.6 It can be obviously seen that theexpected waiting time of a job waiting in a queue is highest for the Separate Queue, sig-nificantly lower in the Class Queue, and lowest in the Common Queue buffering scheme,regardless of policies This is explained by the fact that the Common Queue bufferingscheme allows the most efficient utilization of resource because no job is waiting whenthere is any idle server

For example, with EE policy, at U = 50%, the mean waiting time in Common Queuemodel is 0.00046 (s), while it is 0.00557 (s) in Separate Queue model The difference

of 0.005 (s) is negligible compared to the mean service time of 1.58441 (s) Therefore,

no significant difference is observed concerning the mean response time at U = 50% inFigure 2.5 However,at high system load of U = 90%, the mean waiting time of CommonQueue and Separate Queue model is respectively 0.302311 (s) and 0.693619 (s) The dif-ference is 0.3913 (s), which is clearly observable in Figs 2.5 and 2.6

To quantitatively compare the capability of policies and buffering schemes to fulfilldeadlines (i.e the probability that jobs should be executed within a certain deadline)

we plot the cumulative distribution function (CDF) of response times in Figure 2.7 andFigure 2.8, at 50% and 90% of system load with both policies applied

(a) EE policy applied

0 0.2 0.4 0.6 0.8 1

Class-queue Common-queue Separate-queue

(a) EE policy applied

0 0.2 0.4 0.6 0.8 1

Class-queue Common-queue Separate-queue

(b) HP policy applied

It is observed that the Class Queue and Common Queue buffering schemes have thesame performance concerning the capability to meet deadlines, and they outperforms theSeparate Queue schemes in a certain ranges of deadlines At the medium load, HP policyperforms much better because of prior to high performance processors and CDF reaches

to 1 much faster at limitation value of 6 seconds, while the limitation is the value of 10seconds in case of EE However, at 90% system load when processors in three sites of everymodel are utilized quite equally, there is no noticeable difference between two policies

(a) EE policy applied

1000 1200 1400 1600 1800 2000

(b) HP policy applied

In Figures 2.9 and 2.10 are the mean energy consumption per job in cases of switchingoff and none switching off idle servers, respectively It can be observed that the energyconsumption is independent of the buffering schemes

When idle servers are not switched off (Figure 2.9), the average operating energies,regard to two policies, have the greatest difference noticeable at medium system load,where EE consumes 1769.27 (W.s/job) and HP 1989.82 (W.s/job) This difference de-creases with increasing system load and the energy consumption per job converges to

1550 (W.s/job) at the highest load When idle servers are switched off (Figure 2.10), the

HP and EE policies show opposing tendencies: surprisingly, with HP priority, the energy

EE priority The phenomenon can be explained by the trade-off between performanceand energy saving At medium system load, most job requests are executed on serversoperating with lower energy cost, thus the energy dedicated to the execution of jobs is

significantly smaller (1399 (W.s/job) vs 1649 (W.s/job)) As demand for jobs grows,the use of high performance servers becomes inevitable, since the policies are based onshortest queue In consequence, the difference in AE between SQEE and SQHP decreases

at high system load (1510 (W.s/job) vs 1535 (W.s/job) at 90% of utilization)

The impact on saving energy of switching off servers is illustrated in Figure 2.11 It

is observed that the saving is decreased as the load increases It should also be noticedthat applying switching off idle server technique must take additional cost of recovery andtransition time intervals

(a) EE policy applied

1000 1200 1400 1600 1800 2000

50% 60% 70% 80% 90%

U

No switch-off Switch-off

(b) HP policy applied

We have showed the system performance and energy consumption regarding two ways

of prioritizing servers, high performance and high energy efficiency, in Figures 2.4-2.11.Though servers can be ordered in other ways, we point out that high performance andenergy efficiency are the most promising characteristics to be used for scheduling

We use the Common-Queue model and compare energy efficiency (EE) based, highperformance (HP) based, Low-performance based, Low-power, and Random allocation

in Figures 2.12-2.13 With Low-performance allocation, a free server with lowest mance capacity is selected first In Low-power allocation, a server with lowest active powerconsumption has highest priority Random allocation is the case where the scheduler doesnot care about servers’ characteristics and picks up a free server randomly

0 0.5 1 1.5 2 2.5 3

λ

EE policy Low-Performance

Low-Power Random

0 500 1000 1500 2000 2500

λ

EE policy Low-Performance

Low-Power Random

In Figure 2.12, we can see all these cases perform worse than HP and EE policy,wherein Low-Performance allocation provides worst performance Obviously, HP policy

is the superior to meet the customer’s requirement of service quality In case HP policycannot satisfy the customers’ requirement, the operator should consider to add or replacewith more powerful machines or to restrict the load at a satisfactory level (for instance,the system might block new incoming jobs when the load reaches an expected threshold)

As observed, the effectiveness of HP policy decreases when the system load increases.Figure 2.13 emphasizes the energy savings is achieved most with EE policy Low-powerbased allocation can achieve almost as low energy consumption as EE policy, but results

in worse service time Random allocation and Low-Performance consume more energythan EE and HP policy

To estimate the impact of DVFS, we create a scenario (applying for Common Queuebuffering scheme) where processors reduce their frequency and voltage compared to thefull power (at 100% target load) In this scenario, the ssj_ops value (corresponds to 70%

of the full capacity) and the power consumption of the servers applying DVFS when theyexecute job are as follows:

• Acer AW2000h-Aw170h F2 (Intel Xeon E5-2670): the ssj_ops value is 4517449 andthe active power is 1169(W);

• Acer AW2000h-AW170h F2 (Intel Xeon E5-2660): the ssj_ops value is 3706521and the active power is 881 (W);

• PowerEdge R820 (Intel Xeon E5-4650L): the ssj_ops value is 1961157 and theactive power is 317 (W)

We implement and examine this technique following two approaches:

• Keep the system load ratios by reducing the arrival rates approximately (examinewith range of 50% to 90% of system load)

• Keep the arrival rates system at full workload (examine the two lowest arrival rates

to ensure the stationary property)

In Figures 2.14 and 2.15, we plot the average response time and the average energyconsumption per job vs load for the DVFS scenario and the scenario with the full pro-cessing capacity (denoted as “no DVFS") It worth emphasizing that, at the same loadvalue, the number of arriving jobs is less in the DVFS scenario than in the scenario withthe full processing capacity, and DVFS and “no DVFS" also switch off idle servers Theimpact of DVFS is quite drastic: the increase in the average response time is much higherthan the reduction in the average energy consumption per job

(b) HP policy and Common Queue applied

(b) HP policy and Common Queue applied

In Figures 2.16, 2.17 and 2.18, we plot results when DVFS and “no DVFS" handlethe same number of jobs by keeping the same arrival rate As predicted, the price ofDVFS is the degradation of performance in comparison to full workload operating, which

is clearly observed in Figures 2.16 and 2.17 However, the impact of DVFS on the energyconsumption is only achieved (5.5% energy reduction) in the scenario of the HP policyand Common Queue buffering, whereas it behaves worse than “no DVFS" (4.3% energyincrease) if EE policy is applied instead, (see Figure 2.18)

The results illustrate that DVFS should be carefully tuned if one wants to apply DVFS

to save energy comsumed in computational clusters of heterogenous resources

(b) HP policy and Common Queue applied

No DVFS,lambda = 10.8 DVFS+Switch-off,lambda = 10.8

(a) EE policy and Common Queue applied

0 0.2 0.4 0.6 0.8 1

x (seconds)

No DVFS,lambda = 9.0 DVFS+Switch-off,lambda = 9.0

No DVFS,lambda = 10.8 DVFS+Switch-off,lambda = 10.8

(b) HP policy and Common Queue applied

(b) HP policy and Common Queue applied

This chapter presented a generalized model for the performance evaluation of schedulingcompute-intensive jobs with unknown service times in computational clusters We in-vestigated a set of scheduling schemes that consider the organization of buffers (SeparateQueue, Class Queue and Common Queue) and the resource characteristics of performancecapacity and energy efficiency We defined a ranking methodology of physical servers,which is used to schedule jobs

Numerical results show that the buffering schemes do not affect the energy consumption

of the investigated clusters However, they have a significant impact on the mean responsetime and mean waiting time of incoming jobs Furthermore, the Common Queue and theClass Queue schemes markedly outperform the Separate Queue scheme Therefore, agood buffering scheme can improve the overall cluster performance without an increase

in energy consumption Furthermore, Dynamic Voltage and Frequency Scaling should becarefully applied to reduce the energy consumption of computational clusters