• Tidak ada hasil yang ditemukan

Towards Cost-aware Resource Management in Federated Cloud Data Centers

N/A
N/A
Protected

Academic year: 2023

Membagikan "Towards Cost-aware Resource Management in Federated Cloud Data Centers"

Copied!
178
0
0

Teks penuh

In addition, the geographical diversity of data centers in the federation presents many advantages [9]. Transferring data through data centers significantly increases the TCO of a cloud provider.

Motivation for the Research Work

Optimizing the cost of communication between data centers: Another important factor that affects the TCO of a cloud provider is the cost of communication. In the literature, most works either minimize the request delay or the cost of the cloud provider.

Contributions of the Thesis

The cloud provider must minimize the delay for DS applications, and the delay should not exceed a threshold. FVCC exploits the advantages of the federation for hosting vehicular applications, such as minimizing the cost and delay.

Organization of the Thesis

We present a two-phase algorithm for deploying data-intensive applications in a federated cloud system and then discuss the results from evaluating the proposed algorithm for different scenarios. Finally, we discuss experimental results to demonstrate the efficiency of the proposed algorithm in terms of TCO and migration performance.

Federated Cloud Data Center Architecture

First, we present the general architecture of federated cloud data centers considered in the work and also discuss the vehicular cloud computing map. Federated cloud systems aggregate the resources provided by the data centers of cooperating cloud providers into a single platform [7, 6].

Vehicular Cloud Computing

Vehicles communicate with each other and with RSUs through wireless channels, while RSUs are connected to the central cloud via the Internet. To request a service from the cloud, a vehicle sends a request to the nearest RSU, which forwards it to the CSP.

Virtualization and Migration

When the iterative pre-copy cycle meets one of the defined termination conditions, the migrated VM is suspended in the source data center and the remaining dirty data is fully transferred to the destination data center. So it is the total time to transfer the VM from source to destination.

Taxonomy of Resource Management Problem

Resource Management

Related Work

We then discuss the problem of energy-conscious resource allocation within a single data center and in large-scale data centers. However, they did not account for communication between data centers and the cost of VM migration. [99] looked at the relationship between the response time and the number of selected data centers.

The authors have modeled the response time as the sum of the latency and the processing time of the data center.

Summary

The large resources of a data-intensive application use large amounts of energy, resulting in large electricity bills. Thus, the cost of communication between data centers is another critical cost component in data-intensive applications, where it can exceed the cost of energy consumption. However, the cloud provider should minimize the energy consumption and bandwidth cost together while allocating the VCs for a data-intensive application to reduce its TCO.

Using this framework, we solve the costing problem of data-intensive applications in federated clouds.

Hierarchical Management Framework for Federated Clouds

The FCM of each CP in the federated cloud is responsible for communication with (i) FCMs of other CPs in the federation, (ii) DCMs of the DCs operated by its CP, and (iii) the end users of the associated CPs as well . Since any FCM in the federation can allocate the DC resources, an FCM must request the status information of a particular DC whenever it needs to allocate a workload to it. Upon receiving a request for resource allocation, the FCM collects the status of all DCs across different CPs in the federation by contacting the DCMs (for its own DCs) and the FCMs (for other CPs).

Since status messages are only sent on demand, the framework reduces communication overhead.

System Model and Problem Formulation

F Set of federated cloud administrators in the federation D Set of data centers in the federation. Dp Set of the data centers for thepth cloud provider Dpk kth data center from cloud providerp. Epk Electricity price at data center Dpk (in $/KWh) P U Epk Efficiency of power consumption at data center Dpk spk The number of servers at data center Dpk Ppk Data center power consumption Dpk (KWh) DPpk,ql Data transfer price between Dpk and Dql ($/GB) BWpk,ql The bandwidth of connection between Dpk and Dql (Gbps) C Set of virtual components.

Each cloud provider in the federation has a set of data centers Dp and an FCM Fp.

Proposed Algorithm

Based on this, the algorithm considers the application's traffic pattern and allocates clusters of virtual components instead of individual components. The algorithm iteratively (lines 5-8) removes from C all edges with weight less than or equal to a threshold until C is disconnected (nc>1). Complexity analysis: Let the number of data centers in the federation be M and the number of virtual components in the request beN.

In terms of the required memory, the algorithm should store DClist, the list of data centers in the federation, and the request graph with the set of virtual components and the traffic matrix.

Numerical Results

Federated Cloud

Summary

From the results we conclude that CAVCA is cost-efficient in terms of runtime for large applications and therefore, can be used for on-demand data-intensive application deployment in federal cloud data centers. We proposed a hierarchical resource management framework that meets the requirements of the federated cloud in terms of semi-autonomy and privacy. We formulated the virtual component allocation problem in a federated cloud as an optimization problem to minimize energy consumption and communication costs.

In this chapter, we address the problem of resource allocation for vehicle applications in cloud data centers, where application latency requirements are taken into account while minimizing costs.

System Model

To get a service from the cloud, a vehicle sends a request to the registered service provider through the nearest RSU [137]; requests follow the Poisson distribution [138]. Upon receiving a request, the RSU forwards it to the FCM of the service provider in the cloud after adding RSUid, the forwarding RSU's identifier, as shown in Figure. If an RSU is in the transmission range of the vehicle, the vehicle sends the request directly to the RSU (e.g. V1 .. 4.1), otherwise the vehicle uses multi-hop vehicle-to-vehicle (V2V) communication (e.g.

As resource availability decreases, the fedp rice should increase towards the price offered to the end user.

Proposed Algorithm

The margin between the operating costs and edP rice depends on the resource utilization and it is defined as a percentage of the profit returned from the end user. Therefore f edP ricepkg -the price of a VM instance of type g atDpk offered by CP p to other CPs is given by. Dp Set of the data centers of the CPp Dpk kth data center of CPp.

CDRAM

Simulation Results

It is also observed that CDRAM-DT gives higher latency than CDRAM-DS and DA because it tries to reduce the service cost while fulfilling T hDT. CA gives the highest latency as it only cares about cost minimization and not latency. We can see that the cost with CDRAM-DT is minimal as it distributes and migrates VMs to cheaper DCs that satisfy hDT.

Impact of delay threshold: To investigate the impact of the delay threshold on the cost and the queries served, we fixed the vehicle density at 50 vehicles/km and varied hDS between 100 and 50 ms and dT hDT between 400 and 150 ms.

Summary

We propose an algorithm to migrate VMs between DCs to maximize CP profit while considering migration cost and time. Most works that have addressed the live VM migration problem have also assumed that the VMs are homogeneous. In this chapter, we address the problem of migrating VMs between DCs in a federated cloud to reduce CP TCO.

Based on the proposed cost models, we formulate the VM migration problem between DCs as an optimization problem with the objective of minimizing TCO and migration time.

System Model and Problem Formulation

Based on the stop condition defined in the Xen platform migration algorithm [20], we set the termination condition using three criteria: (i) the number of iterations in the pre-copy stage reached 29, (ii) less than 500 MB was modified during the last pre-copy iteration and (iii) more than three times the size of the VM has been transferred to the destination DC. We express the total size of data transmitted (in GB) during migration of V Mi from. Therefore, we express the cost of inter-DC migration of V Mi from Dpk to Dql (in $) as.

We define T V Cp, the total VM cost of the pth CP, as the operational cost of all VMs served by p.

Proposed Algorithm

If a suitable DesDC is found that makes SrcV M migration favorable with a relatively short migration time, the algorithm migrates SrcV M from SrcDC to DesDC and then updates the location of SrcV M in ResM ap (lines 13-15). The algorithm defines DClist as a list of DCs in the connection, initially including only local DCsDp, and initializes DesDC to N U LL (lines 1-3). Otherwise, the algorithm examines the sorted list of DCs and selects DesDC as the first DC with (i) a positive migration benefit, (ii) has enough resources for SrcV M, and (iii) the available bandwidth between SrcDC and this DesDC is greater than the PDR of SrcV M, which should reduce migration time and downtime (lines 14-18).

The algorithm returns N U LL if there is no DC that satisfies the previous conditions (line 19).

Simulation Results

While setting the destination DC, it considers the network bandwidth and data transfer price to minimize migration time, downtime and migration cost. This directly affects the migration performance and we observe that CAVMA-Fed gives lower migration time and downtime as seen from Figure. CAVMA not only takes into account the smallest price when choosing the destination DC, but also takes into account the connection bandwidth and the price of data transfer between DCs, thus reducing the migration time and cost.

As the PDR gets closer to bandwidth, performance decreases and migration time and downtime increase significantly.

Summary

Fang, “Revenue maximization for dynamic expansion of geographically distributed cloud data centers,” IEEE Transactions on Cloud Computing, vol. Zeghlache, “A mathematical programming approach to revenue maximization in cloud federations,” IEEE Transactions on Cloud Computing, vol. Chen, “A resource-intensive load balancing method for virtual machine migration in cloud data centers,” IEEE Transactions on Cloud Computing , vol.

Buyya, “STAR: SLA-Aware Autonomous Cloud Resource Management,” IEEE Transactions on Cloud Computing, vol.

Referensi

Dokumen terkait

Conclusion of hypothesis test results hypothesis Path Coefficient P values Results Connection Use of Social Media- MSME Performance 0.579 0.000 supported Positive Significant Use