CC BY
111
MODELING OF RESOURCE ALLOCATION IN CLOUD DATA CENTERS
Andrew V. Tutov,
Moscow Technical University of Communications and Informatics, Moscow, Russia, andrew_vidnoe@mail.ru
Natalia V. Tutova,
Moscow Technical University of Communications and Informatics, Moscow, Russia, e-natasha@mail.ru
Anatoly S. Vorozhtsov,
Moscow Technical University of Communications and Informatics, Moscow, Russia, as.vorojcov@mail.ru
Keywords: doud computing, virtual machine allocation, CloudSim, cloud data center, simulation, group method of data handling, GMDH.
Modern commercial data centers include hundreds and thousands of physical servers that consume a significant amount of power. The complexity of modern data center architectures makes it difficult to develop adequate analytical models. Thus the simulation modeling is required. The process of dynamic virtual machine (VM) allocation includes the following steps: monitoring servers to determine the moment when VM migration is needed; selecting VM(s) for migration; the determination of physical servers to host VM(s) based on such criteria as power consumption, temperature, resource utilization and service level agreements. The first stage involve the forecast of changing server parameters in the nearest future. As a method of prediction it is proposed to use the group method of data handling (GMDH). It was noted that the choice of the model structure criterion strongly influences the prediction quality.
The best results were obtained with the model selection by minimizing the criterion of regularity of the last two points of the examination sample. This allows to pay more attention to the latest sample data for a more accurate short-term forecast. To test the effectiveness of the dynamic allocation of resources with the proposed forecasting method a simulation model was developed using the CloudSim package. It was shown that results obtained by the combinatorial GMDH-algorithm in all studied samples are not worse than previously published, obtained by the local regression method. This fact stimulates further research into finding more efficient GMDH-algorithm to achieve the discussed goals. Preliminary analysis of trends in the characteristics of destination hosts on exam subsamples allows us to make an additional condition for selecting the best trend associated with the speed of its change on the sample, which will significantly increase the stability of the migration process of virtual machines.
Information about authors:
Andrew V. Tutov, graduate of the Moscow Technical University of Communications and Informatics, Moscow, Russia.
Natalia V. Tutova, associate professor of the Moscow Technical University of Communications and Informatics, Moscow, Russia.
Anatoly S. Vorozhtsov, associate professor of the Moscow Technical University of Communications and Informatics, Moscow, Russia.
Для цитирования:
Тутов А.В., Тутова Н.В., Ворожцов А.С. Моделирование процессов распределения ресурсов в облачных центрах обработки данных // T-Comm: Телекоммуникации и транспорт. 2017. Том 11. №4. С. 76-80.
For citation:
Tutov A.V., Tutova N.V., Vorozhtsov A.S. (2017). Modeling of resource allocation in cloud data centers. T-Comm, vol. 11, no.4, рр. 76-80.
Introduction
Modern commercial dal ace titers include hundreds and thousands of physical servers that consume a significant amount of power. This number is constantly growing. The adopted amendments to the anti-terror law requiring telecom operators to store up subscribers' traffic for half a year also lead to the increase in the number of servers and virtual machines in data centers. Limited areas of machine halls, high energy costs and the need to meet service level agreements (SLA) require efficient management of data center computing resources and the development of appropriate models and methods. These models should lake into account the capabilities of modern virtualization platforms, such as dynamic scaling of virtual machines and their migration from one physical server to another. Therefore, the theme of this work is relevant.
The complexity of modern data ccnter architectures makes it difficult to develop adequate analytical models. To evaluate the performance and scalability of such systems, as well as test algorithms for resource usage, the simulation modeling tools are required. The advantage of simulation is also in capability of conducting controlled and repeatable experiments to test services under different scenarios of provisioning computing resources.
Related Work
Modeling of dynamic resource allocation in cloud data centers has been addressed in a significant: number of articles. Most of them are private models. Among the works in which the methodology of resource allocation has been introduced are 1131. Algorithms of virtuaI machine allocation with the energy efficiency criterion were proposed in [I], simulations were carried out and the expediency of using the local regression method for forecasting the state of the host in the nearest sliding window was justified. A multi-criteria approach to VM distribution in cloud data centers was proposed in [2]. It was shown thai taking into account several criteria, such as energy efficiency, heat dissipation and resource competency, is more effective, as demonstrated by the prototype of a system using open source software. As a method for predicting the behavior of servers at the next step of monitoring, the method of least squares with the search for parameters in a linear model is proposed.
In work [3] the authors considered the method of multi-criteria virtual machine placement, consisting of three stages: initial, sialic and dynamic allocation. To predict the state of the host during the dynamic allocation process the group method of dala handling {GMDH) is proposed. To test the effectiveness of the proposed forecasting method, simulation modeling has been carried out, the results of simulation are described in this work.
Dynamic resource allocation process
The process of dynamic resource allocation in dala centers includes the following steps:
1. The stage of monitoring servers to determine the level of their workload and the moment when migration of virtual machines is needed;
2. Selecting virtual machines for migration;
3. The determination of physical servers to host virtual machines based on such criteria as power consumption, temperature. the uniformity of resource utilization and satisfying service level agreements.
The first stage involve the determining process based on the forecast of changing server parameters for the nearest period of time. Accurate and fast prediction of server load and temperature at the next step of monitoring will reduce the number of unnecessary migrations and balance the state data center as a whole. As a method of prediction it is proposed to use the group method of data handling (GMDH) [4], which has several advantages such as the possibility of forecasting in terms of statistically insignificant samples, selecting the best model structure among polynomials of given order, test the model in two samples and calculating several types of errors.
To collect time-varying data, such as temperature and CPU utilization, the sliding window method is used. In ease of CPU load or CPU temperature prediction outside the nominal values, the controller initiates the migration process.
In the combinatorial GMDII algorithm, the entire sample A' is divided into three parts: the training Subsample NA, the test sub-sample A'/» and the exam subs ample ;Vr.
n=na+nb+nc.
The training subsample is used to derive estimates for the coefficients of the polynomial, and the test subsample is used to choose the structure of the optimal model. As a selection criterion of the model structure, various criteria can be used [4J.
In this work, we used the minimum regularity criterion (AR)\ ] Nr
AR(B) = — \
Nb t=i
where y,— tabular values of the output variable; y,{B) - values calculated with this model.
In the process of iterating the models by the criterion of regularity, several best models are sclccicd. For the final choice of the model structure, the examination subsample data are taken into account and the regularity criterion is calculated as
, Nc
AR(C) = —ZU--^(C))-
% f=l
min
It was noted that the choice of the model criterion strongly influences the prediction quality. The best results were obtained with the model selection by minimizing the criterion of regularity of the examination sample, making up the last two points. This allows to pay more attention to the latest sample data for a more accurate short-term forecast.
What virtual machines to move depends on the nature of the event.
1. Overheating — VM(s) with the maximum CPU usage have to be migrated, as having the direct impact on the CPU temperature of the physical server. Since the migration process causes extra strain on physical resources, the VM with minimum amount of RAM should be chosen to make the migration faster.
2. The lack of resources - the total average utilization is calculated on the affected physical server and VM are selected from the above-average performance. As in the previous case, all VMs on the server should be sorted in order of increasing size of memory.
3. Low energy efficiency - in this case, the physical server has low resource utilization. All VMs from this server should be moved and the server itself should be switched off.
The next challenge in the dynamic VM allocation process is the determination of the destination host. It should take into ac-
count all the criteria and make the prediction of server slate after migrating VM on it.
Results for various server selection criteria can differ from each other, (-'or this reason it is proposed to use the convolution of multiple criteria into a single objective function having the form as in [5].
Also, before the migration process is started, the state of the destination host should be taken into account to ensure its stability in the near future. For prediction of the server's step-ahead parameters on this stage we also propose to use GMDH-method.
Selecting criteria of I lie best process
Satisfying the requirements on the quality of service is very important for cloud data centers. Quality demands are formalized in the form of SLA-agreements, which can be defined in terms of mctrics such as minimum throughput or maximum response time of the system. Since these indicators can change for a variety of applications, it is important to define application-independent metrics that can be used to estimate the quality of service for any VM in the laaS infrastructure. For the best resource allocation process the following criteria are known:
1. In terms of violation of the SLA-agreements (SLAV, SLA Violation) 111:
SLAV = OTF' PDM
where OTF (Overload Time Fraction) - the period of time during which active hosts are experiencing 100%-utilization.
PDM (Performance Degradation due to Migrations) — overall performance degradation of virtual machines due to migration.
1 s T i W C,
OTF = — Y — > PDM = — V-i-. Nf-J^ M%Crj
where N - the number of hosts; J - the total time during which
si
host i experienced a 100% load that resulted in SLA violations; T - the total time that the host i was in the active state (served
VMs); M — the number of virtual machines; C(/i~ valuation of
the performance degradation of the v irtual machine j in connection due to migrations; C ~ the total CPU performance required
rj
by VM j during its lifetime.
Combined criterion, which include criteria for energy consumption (£) and SLA violations, specified LSV,
ESV -E-SfAV.
2. Total (lata center power consumption in kW*h
3. Thermal efficiency
Maintaining CPU temperature within the specified safe range during operating hours.
4. The stability of the system
It can be characterized by the number of unnecessary migrations and on/off servers.
Because of these changes, it would seem that appropriate VM migration solutions can quickly become irrelevant in terms of performance, temperature or power consumption of destination hosts, which in turn will generate more unnecessary and expensive migrations. Even worse, the system state can fluctuate due
to the continuous migration of VMs, which cause migration and / or action to enable / disable servers. Decisions about when and where to move a virtual machine or when to turn on / off servers should be based not only on current conditions but also on the desirability of this solution for a certain time in the future. Therefore, it is important to make the forecast for the server state as accurate as possible before and after moving virtual machines.
Simulation of allocation process
There are some specialized software packages to simulate cloud data centers, such as CloudS im, CDOSim, DC Sim [6j. The most well-known cloud data center simulation platform is the CloudSim package, which is a set of core classes written in Java which can be used to expand a cloud computing environment with the desired characteristics [71. CloudSim supports data centers modeling, server visualization with customized virtual machine provisioning policies, network topologies, messaging between applications, different clouds and control of simulation progress.
When developing the model, the user describes the simulated system and the simulation scenario in the form of Java source code, if necessary, components can be modified. After starting the program, all data about the simulated system is transferred to the core of CloudSim, where the simulation is performed.
To test the effectiveness of the dynamic allocation of re* sources with the proposed forecasting method a simulation model was developed using the CloudSim package. This package included the GMDM, which was used to predict the state of the serv er at the next step of monitoring.
The simulation was performed for a data center consisting of 800 physical servers, half of which are HP ProLiant MLI10G4 and the other half are the HP ProLiant MLI 10G5. The CPU capacity was measured in millions of instructions per second (MIPS) - 160 MIPS each core for the HP ProLiant MLI 10G5 server and 2660 MIPS for the IIP ProLiant ML110G5 server. Each server has a bandwidth of 1 Gbit/s. Power consumption of servers at different load levels is given in Table I [IJ.
Table 1
Power consumption of the selected server at different load levels in Wxh
Server 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
HP ProLiant G4 86 89.4 92.6 96 99.5 102 106 108 112 114 117
HP ProLiant G5 93.7 97 101 105 110 116 12! 125 129 133 135
Four types of VMs were used in the simulation process, corresponding to the instances of Amazon: 2500 MIPS, 0.85 Gb RAM: 2000 MIPS, 3.75 Gb RAM: 1000 MIPS, 1.7 Gb of RAM; 500 MIPS, 613 Gb of RAM.
The real workload traces available in CloudSim was used. The data contains a CPU load, collected from more than 1000 virtual machines from the Planet Lab project, during a randomly selected 10 days in the spring of 2010. The interval for measur-
m
ing the load was 5 minutes. The load parameters are given in Table 2.
Table 2
Workload datsi parameters (CPU load)
Date Number of VMs Mean St. dev. Ouartile 1 Median Ouartile 3
03/03/2011 1052 12.31% 17.09% 2% 6% 15%
06/03/2011 898 11,44% 16.83% 2% 5% 13%
09/03/2011 1061 10.70% 15.57% 2% 4% 13%
22/03/2011 1516 9.26% 12.78% 2% 5% 12%
25/03/2011 1078 10.56% 14.14% 2% 6% 14%
03/04/2011 1463 12.39% 16.55% 2% 6% 17%
09/04/2011 1358 11,12% 15.09% 2% 6% 15%
11/04/2011 1233 11.56% 15.07% 2% 6% 16%
12/04/2011 1054 11.54% 15.15% 2% 6% 16%
20/04/2011 1033 10,43% 15.21% 2% 4% 12%
At the beginning of the simulation process, each server freely hosted 1-2 virtual machines. During the simulation, virtual machines were consolidated on a smaller number of servers and unloaded servers were turned off to save power.
In the experiments, the window size varied from 7 to 15 points. When forecasting a host load above 85%, a decision to migrate the virtual machine was made. The size of the training subsample in the experiments was 2/3 of the total. The size of the examination subsample was two last points. The group method of data handling was compared with linear models and parameters of these linear models were estimated by the methods of local regression and least squares. As a policy of choosing virtual machines, the policy "minimum time for migration" was taken. The results of the experiments for the window size of 10 points are given in Table 3.
Table 3
The results of experiments
Polity BSV <*10 Power consumption (KWxh) SLAV (x|0S) OTF PDM Number of migrations <*10;)
Local 0,17 163,15 0,11 0,14 0,08 27,63
regression
Least 0,31 185,17 0,25 0,15 0,17 33,86
squares
GMDH 0,16 165,25 0,1 0.1 0,1 29,04
The received results of simulation process of cloud data center resource allocation using the GMDH arc generally comparable to local regression.
When selecting a destination host, its future state was predicted before the virtual machine was moved and placed on it.
Preference was given to the host with a low rate of trend change over the observed period. This avoids the situation when after moving the virtual machine it needs to be returned back or to search for another destination host. This increases the stability of the migration process of virtual machines.
Conclusion
The use of combinatorial GMDH algorithm for predicting the characteristics of destination hosts in all the studied samples is not worse than previously published, obtained by the local regression method. This fact stimulates further research into finding more efficient GMDH algorithm to achieve the discussed above goals.
Preliminary analysis of trends in the characteristics of destination hosts on exam subsamples allows us to make an additional condition for selecting the best trend associated with the speed of its change on the sample, which will significantly increase the stability of the migration process of virtual machines. 7.
References
1. Beloglazov A. and Buyya R., "Optimal online deterministic algorithms and adaptive heuristics for energy and performance efficient dynamic consolidation of virtual machines in Cioud data centers," Concurrency and Computation: Practice and Experience (CCPE), vol.24, no. 13,2012, pp. 1397-1420.
2. Xu J. and Fortes J. A Multi-objective Approach to Virtual Machine Management in Daiaceliters. Proceedings of the 8th International Conference on Autonomic Computing, ICAC 2011, Karlsruhe, Germany. June 14-18, 2011. pp. 225-234.
3. Vorozhtsov A.S., Tutova N. V.. Tutov A. V. The method of optimal virtual servers allocation in data centers. T-Comm. 2015. Vol.9. No.7. pp. 5-10. (in Russian)
4. Ivahnenko A.C., Stepashko V. S Error controlled modeling. Kiev, 1985, 2l6p. (in Russian).
5. Vorozhtsov A.S., Tutova N.V.. Tutov A.V. Optimal cloud servers placement in data centers. T-Comm. 2015. Vol 9. No.6, pp. 4-8. (in Russian)
6. Calheiros. R.N., Ranjan, R.. Beloglazov, A., Rose, C. A. F. D.. and Buyya, R. CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms. Software: Practice and Experience, 41, 2011, pp. 23-50.
7. Mikhailov P.A., Radchenko G.l. Modeling methods and performance evaluation of clotid systems. Vestnik YuUrGU. Series "Computational mathematics and Informatics", 2014, vol. 3, No. 3, pp. 109-123. (ii7 Russian)
МОДЕЛИРОВАНИЕ ПРОЦЕССОВ РАСПРЕДЕЛЕНИЯ РЕСУРСОВ В ОБЛАЧНЫХ ЦЕНТРАХ ОБРАБОТКИ ДАННЫХ
Тутов Андрей Владимирович, МТУСИ, Москва, Россия, andrew_vidnoe@mail.ru
Тутова Наталья Владимировна, МТУСИ, Москва, Россия, e-natasha@mail.ru Ворожцов Анатолий Сергеевич, МТУСИ, Москва, Россия, as.vorojcov@mail.ru
Дннотация
Современные коммерческие центры обработки данных включают в себя сотни и тысячи физических серверов, потребляющих значительное количество электроэнергии. Сложность архитектуры современных ЦОД затрудняет построение адекватных аналитических моделей, поэтому для проверки алгоритмов динамического распределения ресурсов необходимо разработать имитационные модели. Процесс динамического распределения ресурсов ЦОД включает в себя следующие этапы: этап мониторинга серверов с целью определения степени их загруженности и момента начала миграции виртуальных машин; выбор виртуальных машин для миграции; выбор физических серверов для размещения виртуальных машин с учетом таких критериев, как энергопотребление, температура, равномерность загрузки ресурсов и выполнение SLA-соглашений. Характерной особенностью первого этапа является определение момента миграции виртуальных машин на основе прогноза изменения параметров сервера на ближайший период времени. Точный и быстрый прогноз загрузки и температуры сервера на следующем шаге наблюдения позволит уменьшить число ненужных миграций и сбалансировать состояние ЦОД в целом. В качестве метода прогнозирования предлагается использовать метод группового учета аргументов (МГУА). Было замечено, что критерий выбора оптимальной структуры модели сильно влияет на качество прогнозирования. Наилучшие результаты были получены с выбором модели по минимуму критерия регулярности по двум последним точкам экзаменационной выборки. Это позволяет более полно учесть последние данные выборки для более точного краткосрочного прогноза. Для проверки эффективности динамического распределения ресурсов с предложенным методом прогнозирования была разработана имитационная модель с использованием пакета CloudSim. Показано, что комбинаторный алгоритм МГУА для прогнозирования характеристик хостов назначения на всех исследованных выборках даёт результаты не хуже опубликованных ранее, полученных методом локальной регрессии, что стимулирует проведение дальнейших исследований по поиску более эффективного алгоритма Кроме этого, предварительный анализ трендов изменения характеристик хостов назначения на экзаменационных выборках позволяет внести дополнительное условие для выбора лучшего тренда, связанное со скоростью его изменения на данной выборке, что значительно повысит устойчивость процесса миграции виртуальных машин.
Ключевые слова: облачные вычисления, распределение виртуальных машин, CloudSim, облачный центр обработки данных, имитационное моделирование, метод группового учета аргументов, МГУА.
Литература
1. Beloglazov A. and Buyya R. Optimal online deterministic algorithms and adaptive heuristics for energy and performance eficient dynamic consolidation of virtual machines in Cloud data centers," Concurrency and Computation: Practice and Experience (CCPE), vol. 24, no. 13, pp. 1397-1420, 2012.
2. Xu J. and Fortes J. A Multi-objective Approach to Virtual Machine Management in Datacenters. // Proceedings of the 8th International Conference on Autonomic Computing, ICAC 2011, Karlsruhe, Germany, June 14-18, 2011. P. 225-234.
3. Ворожцов А.С., Тутова Н.В., Тутов А.В. Методика оптимального распределения виртуальных серверов в центрах обработки данных // T-Comm: Телекоммуникации и транспорт. 2015. Том 9. №7. С. 5-10.
4. Ивахненко А.Г., Степашко В.С. Помехоустойчивость моделирования. Киев: Наук. думка, 1985. 216 с.
5. Ворожцов А.С., Тутова Н.В., Тутов А.В. Оптимизация размещения облачных серверов в центрах обработки данных // T-Comm: Телекоммуникации и транспорт. 2015. Том 9. No6. С. 4-8.
6. Calheiros, R. N., Ranjan, R., Beloglazov, A., Rose, C. A. F. D., and Buyya, R. CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms. Software: Practice and Experience, 41, 2011, рр. 23-50.
7. Михайлов П.А., Радченко Г.И. Методы моделирования и оценки производительности облачных систем // Вестник ЮУрГУ. Серия "Вычислительная математика и информатика", 2014, т. 3, № 3. С. 109-123.
Информация об авторах:
Тутов Андрей Владимирович, аспирант МТУСИ, Москва, Россия. Тутова Наталья Владимировна, к.т.н., доцент, МТУСИ, Москва, Россия. Ворожцов Анатолий Сергеевич, к.т.н., доцент, МТУСИ, Москва, Россия.
