Efficient Virtual Machine Resource Management for Media Cloud Computing

Abstract

Virtual Machine (VM) resource management is crucial to satisfy the Quality of Service (QoS) demands of various multimedia services in a media cloud platform. To this end, this paper presents a VM resource allocation model that dynamically and optimally utilizes VM resources to satisfy QoS requirements of media-rich cloud services or applications. It additionally maintains high system utilization by avoiding the over-provisioning of VM resources to services or applications. The objective is to 1) minimize the number of physical machines for cost reduction and energy saving; 2) control the processing delay of media services to improve response time; and 3) achieve load balancing or overall utilization of physical resources. The proposed VM allocation is mapped into the multidimensional bin-packing problem, which is NP-complete. To solve this problem, we have designed a Mixed Integer Linear Programming (MILP) model, as well as heuristics for quantitatively optimizing the VM allocation. The simulation results show that our scheme outperforms the existing VM allocation schemes in a media cloud environment, in terms of cost reduction, response time reduction and QoS guarantee.

1. Introduction

In recent years, multimedia cloud computing [1] is becoming a promising technology to provide a flexible stack of computing, storage and software services in a scalable and virtualized manner for media-rich applications [2][3][4]. In this media cloud environment, the virtualization technology is applied to package the required CPU, memory, GPU (graphics processing unit), storage and network bandwidth resources of servers into virtual machines to manage and provision heterogeneous multimedia services and applications at lower cost with minimal efforts. These services include but are not limited to image/video retrieval, video transcoding, streaming, video rendering, media analytics, sharing and delivery [5][6][7].

Due to the heterogeneity and mobility of the media services and users, media cloud has brought up the need for an efficient VM resource management to satisfy the QoS requirements of the media services, especially when different atomic media services such as streaming service, video transcoding services, rendering services and so on are composed to meet the customer demands [8]. These services have different QoS requirements, and need dynamic VM resource capacity at the run-time [1]. In addition, the processing delay of both the atomic and composite media services at the server side under different network conditions makes it difficult to efficiently manage VM resources, while fulfilling QoS demands [1][9][10].

Although there are many researches going on to study various VM resource management techniques in cloud environment [11]-[22], very few of them are suitable for media cloud enthronement (Section2 provides a detailed survey of these works related to current paper). This is due to the dynamic nature of the multimedia services (atomic and composite) in terms of variable resource requirements at run-time. Currently, there exist few researches [23]-[28] related to VM resource allocation in a multimedia cloud environment. However, most of them do not take into account the composite media service scenario, which can affect the response time as well as the overall utilization of the physical resources. In addition, they do not consider the multiple VM resource dimensions (i.e. CPU, GPU, memory, storage and network bandwidth) in the resource allocation problem.

In this paper, we tackle the aforementioned challenges of VM resource allocation in a media cloud environment. We propose a VM resource allocation model that optimally allocates VM resources to a set of physical machines/servers by considering the dynamic VM resource requirements for atomic and composite media services. It also ensures the minimum QoS requirements of the multimedia services, while maintaining high system utilization by avoiding over provisioning the VM resources for the services. The proposed VM resource allocation model is designed to 1) minimize the number of physical servers/machines for cost reduction and energy savings; 2) control the processing delay of multimedia services to improve response time; and 3) achieve load balancing or overall utilization of physical resources.

To the best of our knowledge, none of the follwing works [23]-[28] so far presented as a comprehensive approach that handles all the above objectives at the same time with regards to the resource allocation for a media cloud platform. We formulate the proposed VM resource allocation problem into the multidimensional bin-packing problem, which is NP-complete [29]. Similar to this problem, we consider each virtual machine as an item, and the dimen- sions as its capacities. The goal is to minimize the number of physical servers to be used to place all Virtual Machines, while considering physical servers’ capacities. To reduce the response time, improve the overall resource utilization and avoid frequent VM migration, we additionally consider three constraints such as processing delay, overall resource utilization and special resource utilization. We further define a mechanism that dynamically determines the optimal threshold value of these constraints. Based on the above considerations, we design a MILP model for optimizing VM allocation into physical servers. We also use various heuristics to generate candidate resource allocation schemes and to choose the best scheme for real VM allocation.

Several experiments were carried out to validate the efficiency of our pro- posed VM resource allocation model in media cloud platform. These experiments were conducted for different request patterns of media services in various environments. We have also compared our proposed algorithm with three other existing algorithms in media cloud platform, which comprised of a load balancing model [28], queuing model [25], and a round-robin allocation [21]. The results include the performance of cost reduction, response time reduction and a QoS guarantee.

The rest of the paper is organized as follows: Section 2 presents the related work. Section 3 describes our VM resource allocation model and heuristics for the current multi-dimensional problem. Section 4 presents experimental results and performance comparisons. Finally, Section 5 concludes the paper.

 

2. Related Work

In recent years, the VM resource allocation in cloud environment has gained significant attention. Many existing efforts [11]-[22] study various VM resource management techniques for cloud resource management. Beloglazov et al [11] investigated energy-aware VM resource allocation heuristics that provides data center resources to client applications in a way that improves the energy efficiency of the data center without violating the negotiated SLAs. Aisopos et al [15] proposed a VM resource allocation model for SaaS cloud providers using fractional knapsack problem that maximizes the service provider’s revenue and the resource utilization under a heavy load. The model focuses on maintaining the maximum resource utilization at the cost of risking potential SLA violations over the pending jobs that will yield the smaller profit for the SaaS cloud provider. Lin et al [17] developed a self-organizing model to manage cloud VM resources without the requirement of a centralized management control. Teng et al [18] presented a resource pricing and equilibrium allocation policy for limited cloud user competition over resources. Van et al [16] presented an autonomic virtual resource management mechanism in a cloud for service hosting platforms. Similarly, Berral et al [21] focused on autonomic energy-aware scheduling in the cloud that dynamically adapts to varying task types and workloads, and even to varying cloud infrastructure. The emphasis there on replacing uncertain information with data mining based algorithms, Game-theory based VM resource management for clouds was also studied in [19][20].

However, the works presented above do not consider media cloud environment scenario, where a short battery lifetime, varying wireless channel conditions, and interaction latency pose major challenges for VM resource allocation. Additionally, these VM resource allocation models are unsuitable for heterogeneous multimedia services (atomic and composite), which are basically dynamic in nature. Currently, there are some research efforts going on [23]-[28] for allocating VM resources in a multimedia cloud environment. Sembiring et al. [23] investigated the relationship between the properties of media tasks and their resource consumption and accordingly propose a dynamic resource allocation solution utilizing machine learning techniques. Nan et al [25] proposed a cost-effective resource allocation optimization approach for multimedia cloud that was based on a queuing model. They considered the data center infrastructure as a node-weighted tree-like graph, and then used the queuing model to capture the relationship between the service response time and the allocated resources. They also studied the resource allocation problem in a single-class service case and a multiple-class service case, respectively. The same authors also present similar approach in [24][26][27].

Wen et al [28] presented an effective load-balancing algorithm for a cloud- based multimedia system, which can allocate and schedule VM resources for different user requests with minimum costs. They considered the network proximity and node server traffic loads while making an allocation. They utilized a round-robin algorithm to find the optimal node servers, which would minimize costs. However, those who work in the multimedia cloud assume that the pool of VM resources is homogeneous and that all multimedia service tasks are atomic; this is not practical. In [6], Miao et al presented a cloud-based free viewpoint video rendering framework. In this architecture, every time, the user requests a new viewpoint, the view will be rendered in the cloud and then sent to the client. The researchers also proposed a resource allocation scheme that jointly considered rendering and rate allocations between the cloud and client to optimize the QoE (Quality of Experience). Here, the resource allocation is restricted to the free viewpoint video rendering framework and cannot be used for the general media cloud environment.

Some research [30][31] focused on VM allocation for virtual desktop cloud (VDC) environment from the thin client perspective. In [30], Calyam et al. presents a utility-directed resource allocation model (U-RAM) that uses an offline benchmarking based combined utility function of CPU, memory, net- work health and thin-client user QoE. This is to dynamically create and place virtual desktops in resource pools at distributed cloud data centers, while optimizing resource allocations. This is accomplished by considering timeliness and coding efficiency as quality dimensions. They also propose an iterative algorithm for U-RAM (an NP-Hard problem) to optimize resource allocation with fast convergence that is based on combined utility functions. However, the U-RAM model cannot be directly applicable in the media cloud environment since it does not consider the composite multimedia service or application. This can affect the response time as well as overall utilization of VM resources.

There are some others works which focus on resource allocation in heterogeneous multimedia services and network. For example, in [38], the authors presented a framework for efficient bandwidth allocation and group cooperation in P2P Live streaming scneario. In [39], the authors developed a fair resource allocaiton algorithm based on game theory for wirless multimedia communications. A non-intrusive and adaptable resource management framework, for multimedia applications, distributed in open and heterogeneous home networks is proposed in [40]

To the best of our knowledge, only a few attempts towards proposing an all-inclusive approach that demonstrates a cost-effective and dynamic VM resource allocation for a media cloud platform, which not only considers atomic as well as composite multimedia services but also heterogeneous net- work conditions along with the overall utilization of physical servers. There- fore, the exploration of such a comprehensive approach is timely and crucial for considering the proliferation of emerging multimedia cloud computing.

 

3. Proposed VM Resource Allocation Model in Media Cloud Systems

3.1 Problem Formulation

The proposed VM resource allocation problem is mapped to the multidimensional bin-packing problem [29], which is NP complete. In this problem, we have to map several items into the smallest number of bins as possible. Here, each item denotes a tuple, which contains its dimensions. In our scenario, we consider each VM as an item and the dimensions like CPU, memory, storage, network bandwidth, and GPU, as its capacities. The target is to find out the minimum number of physical servers to place all the VMs, with respect to physical server’s capacities. To solve this problem, we have designed a MILP model, as well as heuristics to quantitatively optimize the VM allocation. We also consider three additional constraints: the processing delay of media service, resource utilization, and the special resource utilization to reduce the response time, improve the overall resource utilization and avoid frequent VM migration respectively.

The VM resource allocation process can be static or dynamic. In static VM allocation, VM capacities are configured using peak load demands of each media workload. The utilization of the peak load demand ensures that the VM does not overload and stay in the same physical servers during their entire lifetime. However, it leads to idleness due to the variable VM resource demand of media workloads.

In dynamic VM allocation, VM capacities are configured dynamically according to the current media workload demands. However, it may require migrating VMs between physical servers in order to: (i) pull out physical servers from an overloaded state when the sum of VMs capacities mapped to a physical server becomes higher than its capacity; (ii) turn off a physical server when the VMs mapped to it can be moved to other physical servers. In our scenario, we allow the virtual machine capacities to be varied on demand. When new media service joins, the system uses the proposed VM allocation algorithm to find proper physical server. If VM migration is required due to the occurrence of an overloaded state, the system also adopts the same VM allocation algorithm for selecting a new physical server. Now, we present a linear programming model to solve the VM resource allocation problem.

3.2 Linear Programming Model

The input parameters and variables used in the linear programming formulation are presented in Table 1. For an atomic or composite media service I that needs to be allocated, the MILP model is presented in Eq. (1) to (6).

The objective function in (1) aims at minimizing the number of required physical servers. The constraint in (2) guarantees that each virtual machine is mapped to a single physical server. Equation (3) guarantees that the virtual machine demands allocated in each physical server do not overload its capacity. The constraint in (4) guarantees that the processing delay of media service I does not exceed a certain threshold value T. Equation (5) is an optional constraint which can help to improve the overall resource utilization. The optional constraint in (6) can reduce the chance of special resource overload and can potentially balance the special resource utilization among all physical servers. Here special resource means that some media applications may give more importance to only CPU than other resources or combinations of any resources. In this scenario, the constraint in (6) cans address to balance the special resource utilization throughout the physical servers.

Table 1.Parameters and variables for the VM allocation problem

Since the future workload may not be predictable, our objective function in (1) represents the average statistics from time 0 to current time. For the constraints, they should be satisfied at any time when the allocation decision is made. Thus, we did not present the time variable t in those formulas. The optional constraints (5) and (6) are proposed to reduce the searching space for this NP-hard optimization problem. However, the use of these constrains may lead the results to be near-optimal. The definitions and effectiveness of rucpv, Spv and dI are explained as follows:

Definition 1: (media service delay): Given a multimedia service I = { v1, v2, ..., vn }, we discuss three types of service flow. The first case is for the atomic multimedia service case, where the VMs have no intercommunication with each other. In this case, the constraint on dI can be defined as follows:

We apply different threshold values for different Virtual Machines ( Tvi for vi ). For any virtual machine vj that does not have communication with an external server, the corresponding constraint can be removed. The second scenario is for the composite multimedia service with synchronous integration, where the screen update from n Virtual Machines is synchronized first, and is then transmitted to the user. In this case, the constraint on dI can be denoted as follows:

Where min(Tvi ) represents the most strict delay constraint among all of the services. Every virtual machine allocation should meet the requirement of min(Tvi ), since they are synchronized.

The last case is for the composite multimedia service with sequential composition, where the output of the service running on vi is the input of the service running on vi+1 . In this case, the constraint on dI can be specified as follows:

In this case, the delay constraint on allocation is computed by subtracting Tv from Tv+1 .

Definition 2 (Resource utilization and Special Resource condition): Given a media service or application i and a physical machine j , let cij , mij , gij , and bij are the percentages of resource usage regarding CPU, memory, GPU and network bandwidth, respectively. If the application is supported by the VM, then the extra resource consumption of VM, OS and remote server should be also included. Thus, application i and virtual machine v can be used interchangeably. Let fcj , fmj , fgj , and fbj be the percentages of idle CPU, memory, GPU and network bandwidth resources, respectively, on machine j . For any resource, if at least k % of free capacity is reserved to buffer the unexpected workload burst, then it would not be counted in the available idle resources. The exact amount of reserved resource is determined by the cloud resource provider by using long term benchmark or short term workload prediction models [32][33]. The media application can be allocated to that physical machine only if the following condition (10) is met:

After media application i is allocated on physical machine j, the average percentage of free resource apij is defined using (11):

where,

For machine j, the resource utilization condition after allocating application i is denoted as rucij , which is a mean-square value. Using (13), we have defined rucij in (5) as follows:

We assume that the applications running on a single physical machine share all of the resource capacity in a proportional way. More specifically, application i will be allocated on physical machine j, which will get (cij / (1- f c' j))× 100% of the CPU capacity, and same for memory, GPU and bandwidth resoueces.

As the resources may not be treated equally in multimedia system, constraint (6) is created to address this issue. Let us consider an instance where CPU is more important for processing the multimedia tasks, such as face recognition or data analytics. Regarding application i, the superiority of machine j is defined as sij in (6) by using (14):

It turns out that for allocation i the superiority of machine j is the percentage of CPU capacity that has been occupied after application i’s allocation. For other systems where bandwidth or GPU is more important, the definition of sij can be modified accordingly.

3.3. Threshold Determination

In our MILP formulation, we have three types of threshold: network delay T in (4), overall resource condition T1 in (5), and free resource T2 in (6). The delay threshold T is given by the QoS requirement of each application. Certain amount of special resource capacity should be reserved to handle an unexpected resource burst. Thus, T2 will be generated according to the specific QoS requirement, the benchmark or the workload burst prediction regarding each application. The applications that experience frequent special resource bursts may require a small T2 value.

Normally, we use the objective function presented in (1) to determine the optimal value of T1 . However, if the physical servers are not capable of supporting all of the requests, in order to find the optimal value of T1 , we will introduce a new objective function U using (15):

where U is the overall resource utilization; ocj , omj , ogj , and obj are the total amount of CPU, memory, GPU and network bandwidth on physical server pj ; λc , λm , λg and λb denote the pricing schemes for each CPU unit, memory unit, GPU unit and network bandwidth unit, respectively; Cp' is the total cost for maintaining the active physical servers. This new objective function measures the total server resource utilization. For the simplicity of explanation, we describe the threshold selection algorithm with the objective function U . It is exactly same as if the objective function U is replaced by the one that we presented in (1). Suppose that we have a set of applications, a set of physical servers and a heuristic adopting threshold T1 . By varying the value of threshold and by repeatedly using the heuristics, we can get different results regarding the overall resource utilization U.

Let f (T1 ) = U (0 <= T1 <=1) be the evaluation function. Our goal is to find an optimal T1 , which can maximize U. Intuitively, the function f (T1 ) should be nearly unimodal in its domain. Considering the optimal T1 as the standard point, a lower T1 will block proper VM allocation while a higher T1 may cause improper VM allocation. Both cases result in low resource utilization since more physical servers have to be used to handle VM requests.

We now propose an iterative approach to determine the optimal value of T1 . The principle of our approach is very similar to that of Bisection method [34] . We repeatedly use the heuristics to simulate the VM allocation process, where different T1 are used. According to the results of U, we bisects the original interval then selects a subinterval where a T1 maximizing f (T1 ) must lie for further processing. The updating rule is shown in Fig. 1.

Fig. 1.The rules for choosing threshold value (a). The peak is located in the side part (b). The peak is located in the middle part.

In Fig. 1, α and β are the midpoints between the current three points. Fig. 1(a) shows the situation where the peak point is located in the side part. If f (α) > f (Mid ), the current midpoint is set as the new high boundary, while α is set as the new midpoint. Similarly, if f (β) > f (Mid ), the current midpoint is set as the new low boundary, while β is set as the new midpoint. As shown in Fig. 1(b), for f (α) > f (Mid ) and f (β) > f (Mid ), α is set as the new low boundary, while β is set as the new high boundary.

According to the definition domain of T1 , the low boundary is initialized to be T1 = 0 , the low boundary is initialized to be T1 = 1 . Therefore, the medium point between the low boundary and the high boundary is T1 = 0.5 . These three points are iteratively updated until the interval between the low boundary and the high boundary is small enough. It is not necessary to determine T1 for each set of VM request during runtime. Only a periodical determination is required.

3.4. Heuristics to Model the VM resource Allocation

The multidimensional bin-packing problem can also be solved using heuristics. Although heuristic solutions will not guarantee an optimal solution, the required time to obtain a feasible solution is much shorter than MILP. We utilize three heuristics like first-fit decreasing (FFD), best-fit decreasing (BFD), and worst fit-decreasing (WFD) [29] and modify them according to the restrictions of the constraints (eq. (1) - (6)). In that heuristics, a lexicographic order is utilized to sort each VM demand. Following the heuristic definitions, the mapping of each VM will then be performed. In the FFD heuristic, the VM will be mapped to the first physical server with available capacity. In case of a BFD heuristic, it will be mapped to the physical server that leaves the least left over space after the mapping between all available physical servers. And for WFD heuristic, it will be mapped to the physical server that leaves the largest left-over space after the mapping between all available physical servers. We adopt these three heuristics to generate candidate VM allocation schemes, and to choose the best scheme for real VM allocation.

3.4.1. Algorithm Description

In this section, we explain the entire procedure of our proposed allocation algorithm. Algorithm 1 presents the pseudo code of the proposed allocation algorithm.

Algorithm 1.Pseudo code of VM resource allocation algorithm

Step 1: Check whether it is necessary to re-select the threshold value. If it is necessary, then run the threshold selection algorithm; Otherwise, use the previous threshold value.

Step 2: Use the three heuristics FFD, BFD and WFD with the threshold value to generate three allocation schemes.

Step 3: Choose the best allocation scheme and enforce allocation.

Step 4: Update the physical server information for next allocation.

3.4.2. Complexity Analysis

The time complexity of the proposed allocation algorithm is denoted by O(3×n2m+9×log2 100×pn2m ). n denotes the number of service request, m denotes the number of physical servers, and p denotes the probability that the threshold will need to be re-selected. The time complexity of heuristic allocation algorithm is O(3×n2m ), since it calls FD, BFD and WFD functions where each function takes O(nm) to allocate one service request, and O(n2m ) for the entire allocation. The threshold determination algorithm consists of log2 100 rounds, and heuristic allocation algorithm is called three times in each round. Thus, each execution of threshold determination algorithm takes O(9×log2 100×n2m ). As it is executed with probability p, the total complexity of threshold determination part is O(9×log2 100×pn2m ).

 

4. Performance Evaluation

In this section, we have presented simulation setup description and conducted several experiments to validate the efficiency of our proposed VM allocation approach as done in. These experiments were conducted for different cases, such as low/high heterogeneity of media tasks, and a large/small media task set. We compared our proposed algorithm with three existing algorithms: a load balancing model [28], a queuing model[25], and a round-robin allocation [21]. The results include the performance of cost reduction; response time reduction and QoS guarantee.

4.1. Simulation Settings

In order to describe the simulation setup in a clear way, we draw the entire simulation infrastructure and present it in Fig. 2. As we can see from the figure, there are two major simulation components: workload generator and cloud simulator. The workload generator takes responsibility of generating atomic, synchronous and sequential workload for simulation. The two sub-components are designed to generate individual multimedia service workload and to generate delay constraints for three types of service composition, respectively. The cloud simulator receives multimedia service requests from workload generator and creates VMs with pre-configured CPU capability, memory, GPU and bandwidth. The cloud resourcesand network environment are simulated by generating physical machines/servers with identical capacity and network latency matrix indicating the network latency between the physical machines/servers, respectively. Finally, seven different resource allocation schemes, including proposed method are tested. The results such as the number of active physical machine/server and average delay are collected from PM (physical machine) monitor and network monitor.

Fig. 2.Silumation Infrstructure

Table 2 shows simulation parameters, where HI, ACU, AMU, APU, ANU represents heterogeneity index, average CPU utilization(%), average memory utilization(%), average GPU utilization(%), and the average network bandwidth utilization(%) respectively. HI, ACU and AMU were retrieved from the Technical University of Berlin (TU-Berlin) workload [8], which are normally used by researchers and students to execute computational experiments. redTo address multimedia service issues in our simulation, we generate the workload by considering several representative multimedia service cases. The first case is Discrete Cosine Transform/Inverse Discrete Cosine Transform (DCT/IDCT), which is very CPU intensive and mostly used in the MPEG and JPEG encoding/decoding [35][36]. Secondly, image rotation has relatively low demand on CPU capacity, but needs more memory and bandwidth. We also consider video rendering workload withhigh demand on GPU capability. As the size of image/video can be different between the service instances, the workload should not be generated in an identical way even for the same multimedia service case. Using the above simulation parameters and the patterns of multimedia service, we randomly generate multimedia service requests in each workload. Initially, the capacities of physical servers were assumed to be identical. In this simulation, the number of physical server was fixed to 100.

Table 2.Details of workload group

Table 3 specifies the delay settings, where IDC, SDC, IDT and SDT represents the individual delay constraint on atomic media service, the sequential delay constraint on adjacent media services, the individual delay time on a single server, and the sequential delay time on connected servers respectively. The variation of heterogeneity only affects the delay constraint on media services.

Table 3.Details of delay

While allocating the services, two types of allocation schemes are adopted. In large group allocation case, all media service requirements are assumed to be generated and submitted at one time. On the other hand, small group allocation allows only 1-5 media service requests submission. When the current allocation is done, the following 1-5 media service requests will be created. In both cases, each group of composite media services contains 1-5 services, which can be atomic, synchronous or sequential. Theservice composition does not introduce more resource consumption, but can change the delay constraint on the services.

4.2. Experiments

4.2.1. Cost Optimization

Several sets of experiments were conducted in the simulation. Firstly, we adopted different request patterns of media services/applications in the experiment (i.e. large/small media service group at Low/high heterogeneous environment) to measure the cost optimization capability. The number of media requests was fixed to 100.

Fig. 3 shows the simulation results derived from the large task group at low heterogeneity environment. From the results, we have found that our proposed approach and the load balancing method have the same performance. The optimal allocation sequence is applied by the load balancing method as well as our proposed approach. Since the low heterogeneity requests do not overuse any type of resource, the resource utilization condition threshold we used in our proposed approach does not provide further optimization in this environment. The queuing model performs worse than any other solution as it does not consider virtual machine. Thus, each service request must occupy one physical server. The round-robin method randomly chooses a physical server for each request. Since it does not provide any optimization method, we have found that 65 physical servers need to be launched according to the allocation results of round-robin method. redThe original FFD, BFD and WFD are also tested in this simulation environment. The results suggest that those original heuristics considering one dimension (CPU capacity) have inefficient performance while dealing with the cloud resource allocation problem on multimedia service. The original WFD, since it always puts VM on a physical server with highest CPU capacity, activates 100 physical servers to handle 100 service requests. The allocation result of original FFD and that of original BFD cause 42 physical servers and 38 physical servers to be active, respectively.

Fig. 3.Cost optimization in large task group at low heterogeneity environmen

In Fig. 4, we present the results retrieved from the large task group and high heterogeneity environment. Due to the resource utilization condition threshold, our proposed approach outperforms existing algorithms by avoiding overuse of any resource. The performance of load balancing method does not degrade so much, as allocating a large group of requests is easier than allocating several small groups of requests. While allocating a large group of requests the load balancing method can always choose the best allocation among all the possible allocations. As the requests consume more resources, the results of round-robin, original FFD and original BFD increases to 68, 48 and 41 active servers, respectively. The performance of queuing method and that of orignal WFD remains same as what we present in Fig. 3 due to the reasons we explained before.

Fig. 4.Cost optimization in large task group at high heterogeneity environment

It can be seen from Fig. 5 that our proposed also achieves best performance in the small task group and low heterogeneity environment. The advantage of our proposed approach was greatly amplified in the high heterogeneity environment, which is more similar to the real scenario. The results are presented in Fig. 6. Compared with the load balancing method, more than a 10% improvement is achieved from using our proposed approach. Thus, we can conclude that our proposed approach is more suitable for the real multimedia service allocation.

Fig. 5.Cost optimization in small task group at low heterogeneity environment

Fig. 6.Cost optimization in small task group at high heterogeneity environment

4.2.2. Scalability

In order to validate the scalability of our proposed approach, we varied the number of service requests to test the cost in the small task group and a high heterogeneity environment. Fig. 7 shows the results of the scalability test. When the total number of service requests equaled 50, the performance of our proposed approach was slightly better than that of the load balancing method. By increasing the workload, a significant difference can be found in the graph. Compared with the loadbalancing method, our proposed approach demonstrates better scalability. The queuing method and the round-robin method have obvious drawbacks regarding scalability.

Fig. 7.Scalability Test Results

The queuing method can support maximum 100 service requests with 100 physical servers. extcolorredThe similar result is observed from the original WFD algorithm except that more than 100 service requests are actually allocated due to the use of VM. Conversely, the round-robin algorithm cannot allocate more than 200 service requests on 100 physical servers. The original FFD and BFD algorithms perform better than round-robin algorithm, but worse than proposed method and balancing method.

In the previous simulations, we allowed our proposed approach to use the optimal resource utilization condition threshold value. It was also necessary to show the importance of the threshold selection. We adopted a different threshold value in our proposed approach and presented the cost optimization results in Fig. 8. From the graph, we can see that our proposed approach can achieve the best performance when the selection of the threshold value is appropriate. Otherwise, the performance may significantly degrade. The results validate the importance of our threshold selection algorithm.

Fig. 8.Resource utilization condition threshold vs. cost

4.2.3. Response time and QoS Success Rate

In the second set of simulations, we explore the actual response time and QoS success rate in different environments. redSince the original FFD, BFD and WFD algorithms do not consider delay issue as the round-robin algorithm does, we do not present similar results of those algorithms. In Fig. 9 and Fig. 10, we present the outcomes that we recorded in a low heterogeneity environment.

Fig. 9.The response time in low heterogeneity environment

Fig. 10.The QoS success rate in low heterogeneity environment

In Fig. 9, the average delay (response time) of the services achieved by each solution is illustrated. The reason why the queuing and round-robin performs is worse than the load balancing and our proposed approach is because they do not consider the delay optimization during the VM allocation process. The loadbalancing method is able to optimize atomic delay for each service allocation. Our proposed approach performs better than the load balancing approach since we clearly define the delay model for both atomic and composite media services. Thus, we are capable to address the dependency issue among the services and the physical machines.

As shown in Fig. 10, our algorithm also achieves a higher QoS success rate. The results of the queuing model decrease dramatically after 100 service requests, as it can only support 100 services with a 100 physical server. The additional service requests must wait in the queue, which means that there is a QoS violation. The round-robin approach maintains an average 70%-80% success rate with random server selection. The load balancing approach cannot meet the QoS requirement when the numberof service requests exceeds 100. Our approach improves this number to 150. At the beginning stage, the composite QoS requirement can be fulfilled by the load balancing approach, since the physical servers with very low individual delay are selected. However, the dependency of services/servers must be considered when the workload increases to a certain level. At that time, the physical servers with very low individual delay have been occupied. Thus, the load balancing approach is more likely to violate QoS requirement without exploring the dependency of services/servers.

Fig. 11 and Fig. 12 show the results in a high heterogeneity environment. From Fig. 11, we can see that the performance of queuing and of the round-robin do not change, since their allocation results are atomic to the service delay requirement. Therefore, the actual response time remains same, as the services are still allocated to the same physical servers. As can be seen from Fig.12 that the QoS success rate dramatically decreases if we use the queuing allocation or the round-robin allocation. The performance of the load balancing method also has an obvious degradation. Our proposed method shows its superiority, including low response time and high QoS success rate in the high heterogeneity environment. Consequently, we can conclude that our delay model is effective in the composite service allocation.

Fig. 11.The response time in high heterogeneity environment

Fig. 12.The QoS success rate in high heterogeneity environment

 

5. Conclusion

The media cloud is emerging as a remarkable technology that can facilitate effective processing of complex multimedia services and provide QoS provisioning for multimedia service or applications from anywhere, anytime and at any device at lower costs. One major challenge for a cloud provider in such media cloud environment is to find an efficient VM resource allocation model for processing media service tasks. This paper presents a VM resource allocation model that dynamically utilizes VM resources to satisfy QoS requirements of media- rich mobile cloud services or applications. In order to do VM resource allocation effectively, we have presented a MILP model, as well as heuristics. Performance comparisons show that our resource management/allocation approach performs very competitively while satisfying users’ QoS demand. This work does not include QoE, media play-back quality, media service profiling and benchmarking. As for the future works, we would incorporate some of above as a part of the future work. We believe that our proposed allocation approach can adapt such settings.