Chapter 5. Minimization of Average Completion Time and Waiting Time in Cloud Computing Environment

Transcription

1 Chapter 5 Minimization of Average Completion Time and Waiting Time in Cloud Computing

2 Cloud computing is the use of the Internet for the tasks the users performing on their computer. Cloud computing, also known as Internet computing, provides on-demand services which are concerned with shared resources, software, information, and other application specific services. Cloud Service Providers (CSPs) provide the service according to the clients requirement with minimal effort at stipulated time [129]. Cloud computing has also the capability to deliver the services to the users by growing or shrinking in instantaneous requirements [14] [12]. Cloud infrastructure implies a business framework between cloud users and CSPs creating a relationship through cloud brokerage services [130] [59]. Cloud brokers [40], acting as negotiator or mediator, help out the users to maintain all the activities such as specific data centre required, completion time of the respective task unit, number of processing cores, waiting time of the respective user request. Cloud broker also helps the users for self-provisioning of resources. In this chapter, CloudSim simulator has been used to find out the average task completion time with respect to the number of task units. Cloud broker is there to decide the average waiting time in the system and also in queue using queuing model Objective of the Chapter In this chapter, we have discussed the CloudSim simulator, queuing model, virtualization, average waiting time. When a huge number of requests have come, the user requests have to wait in the queue and in the system. In this chapter, we have used CloudSim simulator version 3.0 to find out the average task completion time with respect to the number of task units. In the previous chapter, FCFS and Johnson algorithm have been compared to get better scheduling order that has been utilized in queuing model to minimize waiting times, This scheduling order is determined before entering the processing units and this chapter deals with the task scheduling policies within the Virtual Machines. The task units are scheduled as space-shared or time-shared basis within the VMs using simulation toolkit. The simulation test is aimed at the minimization of average task completion time. However, the previous chapter shows that Johnson algorithm produces better outcome using the queuing model with finite capacity and multiple servers. In this relation, we have taken a batch of a 92

3 finite number of task units to be scheduled instead of an infinite number of task units. That has been implemented using M/M/c/K queuing model to reduce the average waiting time which is our main focus of our thesis. Therefore, this chapter has compared M/M/c and M/M/c/K queuing system to analyze the average waiting time in the queue and also in the system CloudSim CloudSim [30] [42] is a popular simulation toolkit that facilitates the user with seamless modeling, simulation, and designing application environments. This simulation framework is developed in the CLOUDS Laboratory, at the Computer Science and Software Engineering Department of the University of Melbourne. This simulation software is released as open source under the LGPL license (Lesser General Public License) CloudSim Platform The main components of CloudSim those are associated with our experimental analysis are shown in the Figure [5.1]. We have presented the correlation between the entities of CloudSim. Users submit their tasks to the broker and broker acts like a dispatcher between data center and user. Data center is correlated with Cloud Information Service (CIS) that registers each data center entry and discovers the resources. A data center encapsulates a set of hosts on which numbers of Virtual Machines (VM) are scheduled. Host models a physical server. VM Scheduler, implemented by the host, models the scheduling policies which are required for allocating processing cores to VMs. VMs are there to execute the Cloudlets which models the cloud-based application services. Each application service contains a preassigned instruction length. Cloudlet Scheduler implements different policies that determine the share of processing power among cloudlets in a VM [30] [42]. 93

4 Figure [5.1]: Structure of CloudSim Functional Model CloudSim Data Flow In this section, we have briefly discussed our simulation work-flow as shown as figure [5.2], and we are going to describe our sequence diagram stepwise as follows: 94

5 Figure 5.2: Sequence Diagram of the Workflow Step 1 : Cloud user sends the job request to the User Interface Step 2 : User Interface analyzes the request according to the Service Level Agreement (SLA). Step 3 : User Interface assigns the tasks to cloud broker. Cloud broker is also termed as Data Center broker. Step 4 : Cloud broker (Data Center broker) divides tasks into same sized 95

6 cloudlets. Cloudlet is also termed as task units. Step 5 : Cloud broker sends the cloudlets to Virtual Machine Manager (VMM). Step 6 : Each data center entity registers with the Cloud Information Service (CIS) registry. Step 7 : Cloud broker sends the resource request to Cloud Information service (CIS). Then Cloud broker consults the CIS to get hold of the record of resources which can recommend hardware services that match user s infrastructure requirements. Step 8 : The cloud broker gets information about the availability of the data center and resources from the CIS. Step 9 : Virtual machine manager (VMM) creates the virtual machine. Step 10 : Data Center entity invokes UpdateVMProcessing for every host that manages it because concerned VMs manage each task unit. That s why the processes should be updated. Step 11 : At the view point of host, invocation of UpdateVMProcessing triggers an UpdateCloudletProcessing process that orders every Virtual Machine (VM) to update its task unit status (executing, finish, suspend) with the data center entity. Step 12 : VM analyze the approximate execution time and sends to host machine. Step 13 : Host machine analyzes the smallest time to the next event. Step 14 : Data center provides the information about the execution time and different resources such as Operating System, VMM used, RAM size, MIPS, a number of cloudlets, the number of CPU, storage capacity etc. Step 15 : Request for execution of the cloudlet is sent to the virtual machine by VMM. 96

7 Step 16 : Cloudlet is being executed in the VM. Step 17 : VM sends the executed cloudlets to the VMM. Step 18 : After completing the execution, VM releases the resources for further use. Step 19 : CIS updates the registry according to the information sent by the data center. Step 20 : VMM further passes the executed cloudlets to cloud broker. Step 21 : Cloud broker combines all the executed cloudlets together to form the task. Step 22 : Cloud broker sends the completed task to the User Interface. Step 23 : After completion of the task, User Interface can either expire the session or make another renewal request. Step 24 : If the session is expired, then User Interface sends the executed task to the cloud user Simulation The main intend of the simulation is to minimize average completion time of the task units using different scheduling policies. An efficient scheduling strategy must provide less response time. We have focused to find out the minimum average completion time of the submitted task units which result in the performance enhancement of the cloud system. There are two scheduling strategies defined in this section i.e. Time-shared scheduling policy and Space-shared scheduling policy. These scheduling strategies can be applied to both VMs and task units. CloudSim allows two level VM provisioning: firstly, at the host level, and secondly, at the VM level. At the host level, it determines the amount of processing power of each processing 97

8 core required for each VM. And at the second level, each task unit is assigned a fixed amount of processing power by the VM. At each level, space-shared and time-shared policies are implemented in CloudSim. These two policies differ with the performance of the task units or application services. In this section, VM provisioning is presented in the Figures [ ]. In this scenario, a host (having two processing cores) receives request for hosting two VMs in such a way that one VM needs two cores and executes four task units. VM1 is assigned to the task units namely T1, T2, T3and T4, and the second one is to host T5, T6, T7 and T8. SPACE-SHARED POLICY: Step 1: Accepted task units are arranged in the ready queue. Step 2: Find out if there are any free processing cores available or not. Step 3: If available, then task units are put in the execution queue. Step 4: First task unit is assigned to the hosting VM. Step 5: After completion of first task, the next task is assigned. Step 6: If the queue is empty, it checks for the new task. Step 7: Then it repeats from the step 1. Step 8: End Note 1: Note 2: As each VM requires two cores, only one VM can run at a given instance of time. Only after completion of VM1, VM2 can be assigned the cores. Each task unit requires only one core. TIME-SHARED POLICY: Step 1: Accepted task units are arranged in the ready queue. Step 2: All the task units are simultaneously assigned to VM. These task units are dynamically context switched during their life cycle. Step 3: When the queue is empty, it checks for the new task. Step 4: Then it repeats from the step 1. Step 5: End 98

9 Note: Each VM gets a time slice on every individual processing core and the time slices are distributed among the task units. Figure [5.3]: VM in space-shared and Cloudlet in time-shared Figure [5.4]: VM in time-shared and Cloudlet in space-shared In figure [5.3], VMs are allocated on the basis of space-shared policy and within a VM, task units are allocated to the processing cores on time-shared basis. As mentioned before, since each VM requires two cores, only one VM can run at a given instance of time. Only after completion of VM1, VM2 can be assigned the cores (Space-shared). Therefore, during a VM lifespan, all the assigned task units are dynamically context switched during their life cycle (Time-shared). In figure [5.4], VMs are allocated in the basis of time-shared policy and within a VM, task units are allocated to the processing cores in space-shared basis. Each VM gets a time slice on every individual processing core and the time slices are distributed among the task units (Time-shared). As the processing cores are shared, VMs are getting fewer amounts of processing power. But, at any given instance of time, only one task can be allocated to each processing core (Space-shared). 99

10 5.3. Result Analysis In this section, we are going to compare the average completion time of the task units. We have configured our simulation environment [132] as stated above. In first case, VMs are allocated in space-shared policy and task units are in time-shared policy. Secondly, VMs are in time-shared policy and task units are in space-shared policy. After the creation of VMs, the task units are submitted, and VMs and task units are arranged accordingly. The tests were conducted on a 32-bit Intel Core i5 machine having 2.60 GHz and 3 GB RAM running windows 7 Professional and Java Development Kit 8 (32 bit). The main goal of our experiment is to compare the average completion time of the said scheduling scenario and to find out the minimum average completion time. We have used NetBeans IDE 8.1 including Java(TM) SE Runtime and CloudSim Toolkit version 3.0 for simulation purpose. In our experimental set up, this simulation creates a data center Broker dynamically and also subject to other constraints this simulation is done. The test simulation environment for measuring the average task completion time includes data center, VMs, and task units (cloudlets). During simulation, a data center is created, which has the characteristics like x86 of architecture, Linux as operating system, Xen as VMM. The simulation environment consists of two hosts; each host has been modeled to have 1000 MIPS, 2 GB of RAM memory, and 1 TB of storage. VM runs inside a host, sharing the host list (the number of hosts present in the simulation) with other VMs. It processes the cloudlets. This processing happens according to a policy (time-shared or space shared), defined by the cloudlet scheduler. The host can submit cloudlets to the VM to be executed. According to our requirement, the following inputs have been taken to create the VMs and cloudlets. 100

11 VM Specification VM ID : Auto incremented depending on the number of VMs (initially 0). Host ID : 0 and 1 as the number of hosts is 2. (ID of the VM s owner i.e. on which host, VM is created ) Capacity : 1000 (in MIPS) RAM : 512 (in MB) Storage capacity : 1000 (in MB) VMM (Virtual : Xen Machine Monitor) Cloudlet Scheduler : Time-shared or Space-shared Number of VMs : 20 Cloudlet Specification Cloudlet ID : Auto incremented depending on the number of cloudlets (initially 0). Length : 4,00000 (the length or size (in MI) of this cloudlet to be executed) (MI: Millions of Instruction) File size : 300 (the file size (in byte) of this cloudlet) Output size : 300 (the output size (in byte) of this cloudlet) Number of Cloudlets : Varies from 20 to 80. Table [5]: Comparison analysis of average task unit completion time Number of Cloudlets Average task unit completion time VM in space-shared and Cloudlet in time-shared VM in time-shared and Cloudlet in space-shared

12 We have performed the test experiment according to the above simulation scenario and the comparison of average completion time of the task units is presented in Table [5]. Figure [5.5]: Comparison analysis of average task unit completion time The above Figure [5.5], shows the comparison analysis of average completion time of task units using the said strategies i.e. space-shared and time-shared policies. Initially VMs are assigned to the hosts and the incoming cloudlets are scheduled aforesaid policies. The number of cloudlets is generally higher than that of the VMs. Therefore, the number of cloudlets and the scheduling policies of cloudlets within the VM are the considerable issues in our simulation. According to the graph, if the cloudlets or task units are scheduled in space-shared policy, it will show better outcome in comparison to time-shared policy. While increasing the number of cloudlets in both the policies, the average completion time of the task units is going high in time-shared policy in comparison with space-shared. When the time span of the cloudlet completes and the cloudlet is still executing on the VMs, the scheduler compulsorily preempts the cloudlets on the VM and allocates the VMs to the next cloudlet [context switch]. Therefore, increasing number of cloudlets leads too much context 102

13 switch [for Time-shared policy] and this phenomenon increases average completion time causing overall degradation of the system performance Queuing Theory Analysis for Cloud Computing As we discussed in the previous chapter, waiting lines or queues have occurred when service demand goes beyond the capacity of the service provider. The queuing model is basically discussed by specifying [114] [99] arrival and service process, number of servers and the maximum capacity of the system. In this chapter, we concentrated on M/M/c queuing model and M/M/c/K queuing model. Assuming that the requests come to the server at Poisson distribution rate and the process time is taken as exponential distribution. It is also assumed that all the processes are non-pre emptive. In those queuing model, c denotes the number of servers and K identifies the places for maximum capacity. So, (K-c) is the queue capacity. We have considered five places of maximum capacity. 1 According to the Kendal s notation [116], average arrival rate will be, E[ ] where τ = Inter-arrival time and E[τ] is denoted as the average or mean Inter-arrival time. 1 Service rate will be, where S is denoted as service time of the customer and E(S) E( S) identifies as the average service time. To make a stable system, an equilibrium condition is to be maintained where the utilization factor Numerical Analysis In this section, we have briefly discussed numerical analysis using queuing model. We have initially taken the average arrival rate and average service rate as shown in the table [6]. Using M/M/c queuing model and M/M/c/K queuing model, we have a comparison study of waiting time. 103

14 Table 6. Initial Parameter (Average arrival rate and service rate [131] ) µ According to queuing system, it has been denoted average number of customers in the system, average number of customers in the queue, average waiting time in the system, and average waiting time in the queue as Ls, Lq, Ws, and Wq respectively. The tables [7-8] show the comparison study using different queuing model. Table 7: Results with M/M/c Queuing Model Lq Ls Wq Ws λ = λ = λ = Table 8: Results with M/M/c/K Queuing Model Lq Ls Wq Ws λ = λ = λ = According to the numerical results we have discussed the comparison study regarding Lq, Ls, Wq, and Ws. The following Figures [ ] show that the average number of customers and the average waiting time in the queue and in the system can be minimized using M/M/c/K rather than M/M/c model. 104

15 Figure 5.6. : Graph Analyzing Average number of customer in the queue (Lq) Figure 5.7. : Graph Analyzing Average number of customer in the system (Ls) 105

16 Figure 5.8. : Graph Analyzing Average waiting time in the queue (Wq) Figure 5.9. : Graph Analyzing Average waiting time in the system (Ws) 106

17 5.6. Summary Rapid use of Internet over the globe, cloud computing has placed itself in every field of IT industry. Recent efforts to make cloud computing technologies better, which includes executing time. Therefore, we have concentrated on simulation-based approaches which help the cloud developers to debug and do performance analysis of their service delivery policies and also their execution time and average waiting time so that the cloud service providers can provide better quality services with minimum execution time. At the end of our work, we can conclude that our sequence diagram and our simulation results may help to grow in cloud infrastructure in surge of fast-growing use of the Internet among the people. In the next chapter, we have presented minimization of waiting time using Genetic Algorithm and queuing model. 107