a distributed virtual architecture for data centers
TRANSCRIPT
Abstraction Layer Based Distributed Architecture for Virtualized Data
Centers
Cloud Computing, Nice, France, 2015.
Ali Kashif Bashir, Yuichi Ohsita, Masayuki MurataGraduate School of Information Science and Technology,
Osaka [email protected]
Project of National Institute of Information and Communications Technology (NICT), Japan.
Objectives
Motivation for the architecture- Use and allocate resources efficiently. - Taking advantage of data correlation among servers.
Architecture should be - flexible to network changes- Scalable to network expansion and changes
Background: Future Internet Model
Infrastructure Providers
Service Provider
Virtual Network ProviderAssembles VNs from InP
Resources
Virtual Network OperatorInstalls and operates VNs
Service ProviderOffers services through VNs
Infrastructure Providers
Service Provider
Current Internet Model
Future Internet Model
Why Virtualization?
Cloud Applications
Internet Users
Flexible, Adaptable, and Scalable Infra. of the Future Internet
Data Center Networks (DCNs) will act as the backbone network for the future Internet.
Network virtualization is a solution to enhance the performance of DCNs.
[1] N.M.M.K Chowdhury, and R. Boutaba, “Network virtualization: state of the art and research challenges,” IEEE. Comm. Mag., vol. 47, pp. 20-26, 2009.[2] K. Tutschku, T. Zinner, A. Nakao and P. Tran-Gia, “Network virtualization: implementation steps towards the future Internet,” Proc. Elec. Comm. Of the EASST., vol. 17, Kassel, Germany, Mar. 2009. [3] M. F. Bari, R. Boutaba, R. Esteves, L. Z. Granville, M. Podlesny, M. G. Rabbani, Q. Zhang, and M.F. Zhani, “Data center network virtualization: a survey,” IEEE. Comm. Survey & Tutor., vol. 15, pp. 909-928, 2013.
Overview of AL-VC
AL-VC: Abstraction Layer Based Virtual Clusters- VMs are grouped into clusters of service types.
• Motivation: In DCNs, servers providing a particular set of services have a high data correlation [1]. Similarly, VMs hosted by these servers also have a high data correlation.
[1] Y.Zhang, A.J. Su, and G. Jiang, “Evaluating the impact of data center network architectures on application performance in virtualized environments,” Proc. 18th Inter. Work. on. Qual. Of. Serv., pp. 1-5, Beijing, China, June 2010.
AL-VC can be deployed on multiple Infrastructure Service Providers (InPs)
Clusters can be formed over a multiple InP’s.
Resource Allocation of AL-VC in the future Internet
…… ……………….InP-1 InP-NInP-3InP-2
…………..
Interface
Virtual Cluster 1
Deploying Virtual Resources
Virtual Cluster 2
Virtual Cluster n
… …… ………….SP-1 SP-3SP-2 SP-N
Users UsersUsers
VNP
VNO
Virtual Network Requests
Abstraction Layer Abstraction layer is created by logically combining subset of VN
switches with an identifier. - This AL is allocated to a particular group of VMs where they altogether form
a cluster.
Assumptions of the architecture Virtual clusters (VCs) are more desirable than physical data
center because the resource allocation in VCs is easily adjustable to meet the changing needs of the users [1].
Devices in a cluster know the topology of the cluster. Network manager, a central entity, decides the cluster formation
and deletion. Decisions regarding cluster adjustments and operations are made
at its AL. With in an AL, all switches are equally capable to process data. To construct VN, we use a particular type of virtual switches
called optical packet switches (OPSs).
[1] C. Guo, G. Lu, J. H. Wang, S. Yang, C. Kong, P. Sun, W. Wu, and Y.Zhang, “SecondNet: a data center network virtualization architecture with bandwidth guarantees,” Proc. 6th Int. Conf. of CO-Next. ACM., New York, USA, Nov. 2010.
Topology (1/2)
All the servers in the server rack are connected to one TOR switch. VMs are connected to TOR switches with multiple links. Communication among VMs is only via OPSs.
Topology (2/2)
An OPS of an AL has four type of connections as: - with TOR switches, - with VMs of local cluster, - with OPSs of local AL, and - with OPSs of VN that are not part of its local AL.
Construction of Abstraction Layer (AL) (1/2)
Basic Idea: Logically allocating a subset of OPS’s to each group of VMs.
Step 1: VMs connect themselves to the switches based on a criterion. Here we use distance as:
WherePi = probability of selecting switches vs
i
dj = distance of switches from VM
ii
jj
RP d
Let’s assume R consists of set of VMs that has link to switches vs
i є Nv
While there is an edge (u, v) where u є R and v ∉ R
Add v to R
Vertex Cover Problem
Construction of AL (2/2)
Step 2: Candidate Set: VMs send list of switches they connect to the network manager.
Step 3: Final Set: Network manager selects the minimum set of common switches that cover all the VMs.
- We assume a graph G = (V, E) with links li 0, where objective is to find a minimum subset of switches that covers all the VMs. For this, we apply
3. Addressing
These switches will be announced as an AL for a cluster and will be assigned with VCID.
Intra-cluster (SID, and IP address) Inter-cluster (SID, VCID, and IP address).
Trade-off- Selecting minimum switches will helps in aggregating the traffic.
On the other hand, too few number of switches may result in contention.
- Therefore, there is a trade-off.
Features of AL-VC Local Management and Control
- AL can provide control to its clusters Scalability
- Number of clusters, number of VMs in a cluster and number of switches in an AL can be scaled to network requirements.
Network Updates- New servers and VMs can be added or deleted in a cluster with less
time and cost. Efficient Utilization of Network Resources
- Bisection of traffic may result in bandwidth improvement- Efficient routing will decrease latency.
Evaluation
Underlying Topology: FATTREE Comparison: Centralized Scheme Evaluation parameters: Efficiency in terms of cost and time when
recovering from network failure (VM, server).
Number of Servers 96Number of VMs 360Max VM a server can host 10Number of switches in AL 10 % of VM in the clusterNumber of clusters 2, 4, 6, 8, and 10DCN topology FATREEParameters Average time and
Communication Cost
Failure of VM (1/2)
.
NF attributes attNv =((att1, nv
1),(att2,nv2),…..,(attn,nv
n)
NF attributes are calculated by the dissimilarity metrics.
Where:- l is the number of NF attributes- disij
r denotes the dissimilarity of VM i and j related to attl. - 𝛅ij
r expresses the coefficients of the NF attributes of machines i and j.
Detection Steps:1. AL detects the failure of VM. 2. It first request the server to host a new VM.3. If server can not, AL request all servers for NF attributes of their VMS. 4. Finally, it deploy VM on the server whose attributes are most matched to the failed VM
Failure of VM (2/2)
Average time required to replace failed VM Communication cost required to replace failed VM
Scalability: Scalable to increase number of clusters. Refined clustering improves efficiency.
Efficiency: Less time and cost to replace the failed VM
Failure of a server (1/2)
Detection Steps:1.When a VM does not respond, AL contacts its server.2.If server also does not also respond, AL assumes that the servers has failed or removed from the cluster. 3.AL informs to the network manager and asks for the attributes of the failed server and its VMs. 4.After receiving NF attributes, it runs a local VM discovery algorithm to find new hosts for the VMs hosted by the server.In case of resource contention, AL request NM to deploy a new server.
Failure of a server (2/2)
Average time required Communication cost required
Scalability, Efficiency.
Here we assume that failing server was hosting 3 VMs
Note: Too many clusters can make network flat.
Summary& Future Works
Network virtualization is essential for the future Internet. AL-VC,
- VMs are clustered into service types.- Each cluster is assigned with a set of virtual switches called an AL. - This architecture tend to offer several features.- Evaluation: It replaced failed VMs and servers with efficiency.
FUTURE WORKS:-Methodology to decide the appropriate number of switches in abstraction layer considering parameters like bandwidth, latency, etc.