A two-phase virtual machine placement policy for data-intensive applications in cloud

Abstract

Cloud computing and big data are two technologies whose combination can yield many benefits and challenges. One of the most significant challenges is the traffic produced by data-intensive jobs within a data center. A possible way to manage the produced traffic is optimizing the placement of virtual machines (VM) on hosts. However, placing VMs compactly to reduce the communication cost can negatively impact on other aspects such as host utilization and load balancing. In this paper, we aim to make a balance between optimizing the host utilization and the communication cost while considering load balancing. We investigate the VM placement problem by modeling it as the minimum weight K-vertex-connected induced subgraph. We prove the NP-Hardness of the problem and propose a novel two-phase strategy for placing VMs on hosts. At first phase, in order to balance traffic and workload among racks, we rank all racks using a fuzzy inference system and select the best ones based on a linear programming model. At second phase, we introduce a novel greedy algorithm to assign each VM to a host regarding a proposed communication cost metric. We evaluate our approach using CloudSim simulator whose results show our two-phase strategy is able to make a balance between host utilization and network traffic. It keeps more than 80 percent of the traffic rack local while reduces the average network link saturation to almost 40 percent with a low variance. Besides, the number of used hosts increase linearly by increasing the number of VMs which leads to a higher host utilization.

Introduction

Data-intensive applications such as image processing and machine learning, play an increasingly important role in many domains of business and scientific research. They are required to efficiently analyze a huge amount of data which is known as big data. In terms of big data analysis, the data-parallel strategy is an efficient approach to process large amounts of data on a cluster of computing machines. On the other hand, big data processing tools require a host cluster such as private cloud whose maintenance is not feasible for a majority of companies and institutes. Therefore, utilizing a public cloud infrastructure is a common and acceptable solution (Assunção et al., 2014, Philip Chen and Zhang, 2014, Hashem et al., 2015).

As virtual machines running a data-intensive application transfer a large amount of data among each other, running big data analytical jobs on the public cloud highly effects network traffic. Based on Cisco global cloud index 2016–2021, big data as a significant driver of data center traffic will be responsible for 20 percent of all traffic within the data center by 2021 which was 12 percent in 2016. Inefficient management of this traffic results in various undesirable effects like data transferring latency, high job completion times, network failure , etc. (Mann, 2015, Cisco, 2018). In addition, by rising the popularity of some platforms like Kubernetes which uses a virtual cluster of VMs to provide a specific service including data analyzing, the importance of considering the communication between VMs has been increased.

An optimized VM placement can improve the network scalability and reduce the communication cost to a great extent. Nevertheless, a majority of the preposed placement schemas model the VMs as independent objects and assign them to the physical machines without considering consumption of network resources (Filho et al., 2017, Masdari and Zangakani, 2019). On the other hand, the proposed methods for reducing the communication cost either contain some assumptions which make them impractical or do not consider others essential aspects of infrastructure management such as resource utilization and load balancing.

In this work, in contrary to previous works, we do not consider any predefined traffic among virtual machines. The reasons for elimination of this assumption are (1) it is hard to predict traffic among each pair of virtual machines beforehand, (2) in real-world environments these simplifying assumptions do not hold due to the dynamic interactions between these virtual machines. Alternatively, we assume there is a high amount of traffic with an unknown pattern and search for a cluster of close hosts whose connecting links have the maximum available bandwidth. However, this assumption may lead to a suboptimal solution where the communication pattern among VMs is neither many-to-many nor one-to-many. Therefore, the optimal placement is achieved by dividing VMs in groups as a majority of previous works has done.

In addition to communication cost, we consider improving host utilization as the second optimization objective which is a main concern for cloud service providers. In order to address the problem of east–west traffic flow within a data center, we only consider big data processing batch jobs where data is stored inside the VMs and exchanged between them as network traffic.

We also consider a one-to-one mapping for the assignment of the VMs to the hosts. Some of the previous placement schemas try to assign all the VMs to a single host which causes a low degree of reliability in the case of host failure (Liu et al., 2019). To decide about how many VMs of an application are allowed to place on a single host is out of the scope of this research and is discussed in Section 6 as the future work. Therefore in this work, we consider a high degree of reliability which means no two VMs belonging to the same application are placed on the same host.

In summary, in this paper the virtual machine placement (VMP) problem is investigated to improve the data center network performance and resource usage. Our work aims to select a group of hosts to place a cluster of virtual machines on them in order to keep the traffic more local and balanced inside a rack as well as between different racks in a data center while considering the hosts’ utilization. For this reason, we first try to find a specific area of data center with regard to our objectives, and then select the final hosts from the specified area. The target area, in the ideal case, is a single rack with enough resources and low network traffic, and if it is not possible, a combination of a minimum number of racks which are more likely to handle the request efficiently. In this work, the term distance refers to the number of hops between the source and the destination in a data center.

The main contributions of this work are summarized as following:

• A new metric, namely networking-cost, is defined to calculate the communication cost between hosts. It considers both the traffic and the number of hops between hosts.

• A fuzzy inference system (FIS) is designed to rank the racks based on their total free resources and intra-rack traffic.

• A linear programming (LP) model is defined to select a minimum number of close racks having high rank while the traffic on their uplinks is low.

• A greedy method is implemented to assign VMs to the hosts inside the selected rack(s) based on the metric networking-cost in order to create a balanced traffic in the data center.

The remainder of this paper is organized as follows: In Section 2, we review the related works. In Section 3, we introduce our VM placement problem and prove its NP-Hardness characteristic. In Section 4, we present our strategy which solves the problem in two phases. In Section 5, the evaluation results are analyzed; and in the final Section 6, we conclude our research and discuss the possible future works.