基于LSTM和GRNN的容器配额优化算法

doi:10.12263/DZXB.20201211

摘要/Abstract

摘要：

为了实现容器配额设置自动化和集群资源利用最大化，本文设计了一种容器配额优化算法.本文在长短期记忆（Long Short-Term Memory，LSTM）神经网络和广义回归神经网络（Generalized Regression Neural Network，GRNN）的基础上设计了深度神经网络（Long short-term memory and GRNN Network，LGN），并使用改进量子粒子群算法优选网络结构超参数，以实现自动调参和更快的收敛速度.容器配额优化算法步骤如下：首先根据历史数据使用LGN训练资源容量模型，然后使用改进的量子粒子群算法优化模型参数，最后使用资源容量模型计算容器配额.通过与谷歌容器垂直自动扩展器（Vertical Pod Autoscaler，VPA）和水平自动扩展器（Horizontal Pod Autoscaler，HPA）生成的配额进行对比发现，本文提出的优化算法较VPA和HPA降低了至少10%的资源分配总量，同时提升了至少6%的资源利用率.

关键词: 容器配额, 容量模型, 广义回归神经网络, 长短期记忆神经网络, 量子粒子群算法

Abstract:

In order to realize the automation of container quota setting and the maximization of cluster resource utilization, this paper designed a container quota optimization algorithm. In this paper, LGN(long short-term memory and GRNN network) was designed based on the LSTM(long short-term memory) and the GRNN(generalized regression neural network). Also, the improved quantum particle swarm algorithm was used to optimize the hyperparameters of the network structure to achieve automatic parameter adjustment and faster convergence speed. The steps of the container quota optimization algorithm are as follows. First, LGN is used to train the resource capacity model based on historical data. Then the improved quantum particle swarm algorithm is adopted to optimize the model parameters. Finally, the resource capacity model is designed to calculate container quotas. By comparing with the quotas generated by Google Container VPA(vertical pod autoscaler) and HPA(horizontal pod autoscaler), the optimization algorithm proposed in this paper has at least 10% lower total resource allocation and 6% higher resource utilization.

Key words: container quota, capacity model, generalized regression neural network(GRNN), long short-term memory(LSTM), quantum particle swarm optimization

中图分类号:

TP18

周泓岑, 白恒, 才振功, 蔡亮, 顾静, 汤志敏. 基于LSTM和GRNN的容器配额优化算法[J]. 电子学报, 2022, 50(2): 366-373.

ZHOU Hong-cen, BAI Heng, CAI Zhen-gong, CAI Liang, GU Jing, TANG Zhi-min. Container Quota Optimization Algorithm Based on GRNN and LSTM[J]. Acta Electronica Sinica, 2022, 50(2): 366-373.

图/表 14

图1 配额优化算法结构图

图2 广义回归神经网络结构图

表1 配额优化问题定义

服务实例数及约束

$S$ ：服务集合.

$P s$ ：服务 $s$ 的实例集合 $P s$ .

$n s$ ：服务 $s$ 的实例个数，满足约束条件 $n s m i n ≤ n s ≤ n s m a x$ .

容器

尺寸

$R c (p)$ ：实例 $p$ 的CPU需求，满足 $R c m i n (p) ≤ R c (p) ≤ R c m a x (p)$ .

$R m (p)$ ：实例 $p$ 的内存需求，满足 $R m m i n (p) ≤ R m (p) ≤ R m m a x (p)$

机器

尺寸

$M$ ：机器集合.

$C c (m)$ ：机器 $m$ 的CPU容量.

$C m (m)$ ：机器 $m$ 的内存容量.

容量

模型

Q：服务请求量.

$U s c$ ：服务 $s$ 的CPU资源使用量.

$U s m$ ：服务 $s$ 的内存使用量.

$F s c (Q, T)$ ：服务请求量 $Q$ 、时间 $T$ 和CPU资源使用量映射函数，等于容量模型的训练结果.

$F s m (Q, T)$ ：服务请求量 $Q$ 、时间 $T$ 和内存资源使用量映射函数，等于容量模型的训练结果.

表1 配额优化问题定义

服务实例数及约束

$S$ ：服务集合.

$P s$ ：服务 $s$ 的实例集合 $P s$ .

$n s$ ：服务 $s$ 的实例个数，满足约束条件 $n s m i n ≤ n s ≤ n s m a x$ .

容器

尺寸

$R c (p)$ ：实例 $p$ 的CPU需求，满足 $R c m i n (p) ≤ R c (p) ≤ R c m a x (p)$ .

$R m (p)$ ：实例 $p$ 的内存需求，满足 $R m m i n (p) ≤ R m (p) ≤ R m m a x (p)$

机器

尺寸

$M$ ：机器集合.

$C c (m)$ ：机器 $m$ 的CPU容量.

$C m (m)$ ：机器 $m$ 的内存容量.

容量

模型

Q：服务请求量.

$U s c$ ：服务 $s$ 的CPU资源使用量.

$U s m$ ：服务 $s$ 的内存使用量.

$F s c (Q, T)$ ：服务请求量 $Q$ 、时间 $T$ 和CPU资源使用量映射函数，等于容量模型的训练结果.

$F s m (Q, T)$ ：服务请求量 $Q$ 、时间 $T$ 和内存资源使用量映射函数，等于容量模型的训练结果.

表2 数据集格式示意表

时间戳	QPS	内存使用量	CPU使用量
2019/4/1 0：00	52 039	0.892 6	0.78
2019/4/1 0：05	50 301	0.842 4	0.65
2019/4/1 0：10	56 726	0.928 5	0.92

表3 资源使用预测中各种模型的预测误差

模型名	RMSE	MAE	R²
LGN	121.05	78.67	0.92
LSTM	157.13	92.53	0.65
BiLSTM	176.25	105.24	0.74
GRNN	158.39	102.84	0.79
BP	267.40	155.03	0.54
ARIMA	366.84	167.46	0.82
RF	265.23	145.86	0.89

图3 改进量子粒子群算法QPSO和原始量子粒子群算法oQPSO收敛速度对比

图4 5种配额方案在不同负载下CPU配额

图5 5种配额方案在不同负载下内存配额

图6 5种配额方案在不同负载下CPU利用率

图7 5种配额方案在不同负载下内存利用率

表4 同负载下资源配额

实验组	CPU配额	内存配额	实例数
最小配额数方案	0.995 c	1 782 m	3
最大配额数方案	0.298 c	534 m	10
等比例配额方案	0.498 c	891 m	6
VPA配额方案	0.658 c	1 054 m	6
HPA配额方案	1 c	2 000 m	4

表5 同负载下资源分配总量

实验组	CPU总分配	内存总分配
最小配额数方案	2.987 c	5 347 m
最大配额数方案	2.987 c	5 347 m
等比例配额方案	2.987 c	5 347 m
VPA配额方案	3.948 c	6 324 m
HPA配额方案	4 c	8 000 m

表6 同负载下资源利用率

实验组	CPU	内存
最小配额数方案	80.2%	81.6%
最大配额数方案	83.1%	82.7%
等比例配额方案	86.4%	80.5%
VPA配额方案	64.7%	75.2%
HPA配额方案	74.2%	64.3%

图8 HPA方案实例数随负载变化

参考文献 26

1	GOOGLE CLOUD.Vertical pod autoscaling[EB/OL]..
2	JUNG G, SIM K M. Agent-based adaptive resource allocation on the cloud computing environment[C]//2011 40th International Conference on Parallel Processing Workshops. Piscataway， NJ, USA: IEEE, 2011: 345-351.
3	AKHTER N, OTHMAN M. Energy aware resource allocation of cloud data center: Review and open issues[J]. Cluster Computing, 2016, 19(3): 1163-1182.
4	YIN L, LUO J, LUO H. Tasks scheduling and resource allocation in fog computing based on containers for smart manufacturing[J]. IEEE Transactions on Industrial Informatics, 2018, 14(10): 4712-4721.
5	GEETHA P, ROBIN C R R. A comparative-study of load-cloud balancing algorithms in cloud environments[C]//2017 International Conference on Energy, Communication, Data Analytics and Soft Computing(ICECDS). Piscataway, NJ, USA: IEEE, 2017: 806-810.
6	ALAM A B M B, ZULKERNINE M, HAQUE A. A reliability-based resource allocation approach for cloud computing[C]//2017 IEEE 7th International Symposium on Cloud and Service Computing(SC2). Piscataway, NJ, USA: IEEE, 2017: 249-252.
7	周景才, 张沪寅, 查文亮, 等. 云计算环境下基于用户行为特征的资源分配策略[J]. 计算机研究与发展, 2014, 51(5): 1108-1119.
	ZHOU J C, ZHANG H Y, ZHA W L, et al. User-aware resource provision policy for cloud computing[J]. Journal of Computer Research and Development, 2014, 51(5): 1108-1119. (in Chinese)
8	ISLAM S, KEUNG J, LEE K, et al. Empirical prediction models for adaptive resource provisioning in the cloud[J]. Future Generation Computer Systems, 2012, 28(1): 155-162.
9	丁丁, 罗四维, 艾丽华. 基于双向拍卖的适应性云计算资源分配机制[J]. 通信学报, 2012, 33(Z1): 1-143.
	DING D, LUO S W, AI L H. Adaptive double auction mechanism for cloud resource allocation[J]. Journal on Communications, 2012, 33(Z1): 1-143. (in Chinese)
10	BALLA D, SIMON C, MALIOSZ M. Adaptive scaling of Kubernetes pods[C]//NOMS 2020-2020 IEEE/IFIP Network Operations and Management Symposium. Piscataway, NJ, USA: IEEE, 2020: 1-5.
11	RATTIHALLI G, GOVINDARAJU M, LU H, et al. Exploring potential for non-disruptive vertical auto scaling and resource estimation in kubernetes[C]//2019 IEEE 12th International Conference on Cloud Computing(CLOUD). Piscataway, NJ, USA: IEEE, 2019: 33-40.
12	ROSSI F, NARDELLI M, CARDELLINI V. Horizontal and vertical scaling of container-based applications using reinforcement learning[C]//2019 IEEE 12th International Conference on Cloud Computing(CLOUD). Piscataway, NJ, USA: IEEE, 2019: 329-338.
13	AL-MAHASNEH A J, ANAVATTI S G, GARRATT M A, et al. Evolving general regression neural networks using limited incremental evolution for data-driven modeling of non-linear dynamic systems[C]//2018 IEEE Symposium Series on Computational Intelligence(SSCI). Piscataway, NJ, USA: IEEE, 2018: 335-341.
14	SHAO S, CHEN W, CHEN J. A noisy data regression model based on general regression neural networks[C]//2011 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE 2011). Piscataway, NJ,USA: IEEE, 2011: 160-163.
15	GOULERMAS J Y, LIATSIS P, ZENG X, et al. Density-driven generalized regression neural networks(DD-GRNN) for function approximation[J]. IEEE transactions on neural networks, 2007, 18(6): 1683-1696.
16	HOCHREITER S, SCHMIDHUBER J. Long short-term memory[J]. Neural computation, 1997, 9(8): 1735-1780.
17	谢晓兰, 张征征, 郑强清, 等. 基于 APMSSGA-LSTM 的容器云资源预测[J]. 大数据, 2019, 5(6): 62-72.
	XIE X L, ZHANG Z Z, ZHENG Q Q, et al. Container cloud resource prediction based on APMSSGA-LSTM[J]. Big Data Research, 2019, 5(6): 62-72. (in Chinese)
18	KUMAR S D, SUBHA D P. Prediction of depression from EEG signal using long short term memory(LSTM)[C]//2019 3rd International Conference on Trends in Electronics and Informatics(ICOEI). Piscataway, NJ, USA: IEEE, 2019: 1248-1253.
19	ASHWINI B C, NIJAGUNARYA Y S. Resource scheduling in cloud computing using back propagation algorithm[C]//2017 2nd International Conference on Emerging Computation and Information Technologies(ICECIT). Piscataway, NJ, USA: IEEE, 2017: 1-6.
20	KUMAR S, PANDEY M K, NATH A, et al. Missing QoS-values predictions using neural networks for cloud computing environments[C]//2015 International Conference on Computing and Network Communications(CoCoNet). Piscataway, NJ, USA: IEEE, 2015: 414-419.
21	RAO V, RAO S. Application of artificial neural networks in capacity planning of cloud based IT infrastructure[C]// 2012 IEEE International Conference on Cloud Computing in Emerging Markets(CCEM). Piscataway, NJ, USA: IEEE, 2012: 1-4.
22	WILSONPRAKASH S, DEEPALAKSHMI P. Artificial neural network based load balancing on software defined networking[C]//2019 IEEE International Conference on Intelligent Techniques in Control, Optimization and Signal Processing(INCOS). Piscataway, NJ, USA: IEEE, 2019: 1-4.
23	KENNEDY J, EBERHART R. Particle swarm optimization[C]//Proceedings of ICNN'95-International Conference on Neural Networks. Piscataway, NJ, USA: IEEE, 1995: 1942-1948.
24	张利彪, 周春光, 马铭, 等. 基于粒子群算法求解多目标优化问题[J]. 计算机研究与发展, 2004, 41(7): 1286-1291.
	ZHANG L B, ZHOU C G, MA M, et al. Solving multi-objective optimization problems based on particle swarm optimization[J]. Journal of Computer Research and Development, 2004, 41(7): 1286-1291. (in Chinese)
25	DING W, FANG W. Target tracking by sequential random draft particle swarm optimization algorithm[C]//2018 IEEE International Smart Cities Conference(ISC2). Piscataway, NJ, USA: IEEE, 2018: 1-7.
26	PEERTUBE.What is PeerTubef?[EB/OL].(2020-06-26) [2021-12-30]. .