Download presentation
Presentation is loading. Please wait.
Published byMay Dickerson Modified over 9 years ago
1
DESIGN CONSIDERATIONS OF A GEOGRAPHICALLY DISTRIBUTED IAAS CLOUD ARCHITECTURE CS 595 LECTURE 10 3/20/2015
2
“Computation may someday be organized as a public utility.” - John McCarthy, 1961 2
3
OUTLINE Cloud Computing Introduction Local Cloud Architectures (IaaS) Current Research Topics Reducing costs of Owning/Operating Private Clouds Deployment Power Aware Load Consolidation Power Aware Storage Consolidation Virtual Machine Migration Conclusion Future Work 3
4
INTRODUCTION CLOUD DEFINITION Cloud computing is a set of service-oriented architectures, which allow users to access a number of resources in a way that is elastic, cost-efficient, and on-demand. 4
5
INTRODUCTION CLOUD DEFINITION Scalable resource allocation Tailored services Software as a Service (SaaS) Platform as a Service (PaaS) Infrastructure as a Service (IaaS) Billed like a utility public clouds 5
6
INTRODUCTION CLIENT/SERVER VS. CLOUD ARCHITECTURE Network Client Server Network Client Cloud Interface Compute Node Switch/ Router Storage Node Client/Server Architecture Cloud Architecture Storage Client Cloud Admin 6
7
LOCAL CLOUD ARCHITECTURES IAAS Local Cloud? Small to medium sized What resources would we need to do this? Compute Servers Persistent Storage Servers VM Image Server(s) Cloud Administrative Server(s) Network Infrastructure Copper 7
8
LOCAL CLOUD ARCHITECTURES IAAS Compute Servers CPU, RAM, Local disk (magnetic, SSD) resources given to the user. In the form of virtual machines. Hosts virtual machines using a hypervisor Xen, KVM, ESXi Grid VGX Hybrid approach to hypervisor selection is common. Windows Linux Mac OS X 8
9
LOCAL CLOUD ARCHITECTURES IAAS Persistent Storage Servers What they are: VMs hosted on the compute servers are stateless. What they do: Used for long term storage of data. Virtual Machine Image Server Modified Persistent Storage Server. Repository of available VM images. 9
10
CURRENT RESEARCH Private cloud architectures are great! But how can we expand while staying cost efficient? Current Research Areas: Deployment with limited networking resources. Power aware cloud resource control. Power Aware Load Consolidation (PALC) Power Aware Storage Consolidation (PASC) Virtualization: Migration, Performance, and Costs 10
11
CURRENT RESEARCH IAAS ARCHITECTURE 11
12
CURRENT RESEARCH INTERFACE - RELATED Lonea, et al. [2012] Interfaces for Eucalyptus Hashimoto, et al. [2012] User Interface of Overlay Networks for Clouds Elasticfox [2013] Amazon AWS Web Interface, Firefox plugin Hybridfox [2013] Eucalyptus Web Interface, Firefox plugin 12
13
CURRENT RESEARCH INTERFACE Great! Now how do I interface with all of these cloud resources? 13
14
CURRENT RESEARCH INTERFACE 14
15
CURRENT RESEARCH INTERFACE Now that we have the cloud infrastructure, time to start using the resources! One NAT router, one public IP address. Problem? Only the cloud resources are behind the NAT router. Solutions VPN for client devices Extra layer of complexity for user NAT port forwarding Uses DB to maintain protocols currently used by clients Update IP tables Hidden from user 15
16
CURRENT RESEARCH INTERFACE 16
17
CURRENT RESEARCH INTERFACE 17
18
CURRENT RESEARCH INTERFACE 18
19
CURRENT RESEARCH GEOGRAPHICALLY DISTRIBUTED PRIVATE CLOUD ARCHITECTURE - RELATED Nielsen, Hacker 2010 Using VPN for connecting VM based HPC Systems Nimbula 2013 Oracle based cloud infrastructure management Wu, et al. 2011 Plume, distributed cloud network for task dissemination We provide an architecture for users to connect to a geographically distributed private IaaS cloud. 19
20
CURRENT RESEARCH GEOGRAPHICALLY DISTRIBUTED PRIVATE CLOUD ARCHITECTURE 20
21
CURRENT RESEARCH GEOGRAPHICALLY DISTRIBUTED PRIVATE CLOUD ARCHITECTURE 21
22
CURRENT RESEARCH GEOGRAPHICALLY DISTRIBUTED PRIVATE CLOUD ARCHITECTURE Distributed: Main Cloud Cluster Many auxiliary cloud clusters All private, behind NAT routers Connected using EoiP Advantages Pooling of resources Issues? Network traffic between cloud clusters Clients – use which cloud cluster? 22
23
CURRENT RESEARCH GEOGRAPHICALLY DISTRIBUTED PRIVATE CLOUD ARCHITECTURE 23
24
CURRENT RESEARCH GEOGRAPHICALLY DISTRIBUTED PRIVATE CLOUD ARCHITECTURE Example of Availability Zone’s in different cloud clusters: * Eucalyptus 2.1 24
25
CURRENT RESEARCH GEOGRAPHICALLY DISTRIBUTED PRIVATE CLOUD ARCHITECTURE 25
26
CURRENT RESEARCH POWER CONSUMPTION Reducing costs by applying active power aware strategies. Power Aware Load Consolidation Power Aware Storage Consolidation Advantages lowers operational costs of cloud resources Issues Resource availability power state transitions 26
27
CURRENT RESEARCH POWER CONSUMPTION – PALC - RELATED Van, et al. 2011 Power vs Performance in Clouds Barroso, Holzle 2012 Energy-Proportional Computing Hu, et al. 2011 Scheduler for Load Balancing in Cloud Computing We provide a power aware strategy for consolidating virtual machine requests on compute nodes. Power down unused compute nodes 27
28
CURRENT RESEARCH POWER CONSUMPTION - PALC PALC Placement of virtual machines on as few compute servers as possible. Other compute servers in low power state Able to convert hot/cold depending on user demand 28
29
CURRENT RESEARCH POWER CONSUMPTION - PALC 29
30
CURRENT RESEARCH POWER CONSUMPTION - PALC Algorithm PALC consolidate: for all active compute nodes j ∈ [m] do n j current utilization of compute node j end for if all n j > D t utilization //all available nodes are active boot vm on most underutilized compute node end if else boot vm on most utilized compute node end else upscale: if for all n > D t utilization if number of active compute nodes < m boot next available compute node end if downscale: if vm i idle > 6 hours or user initiated shutdown shutdown vm i end if if n j has no active vm shutdown active compute node end if 30
31
CURRENT RESEARCH POWER CONSUMPTION - PALC VM request schedule for PALC experiments: 31
32
CURRENT RESEARCH POWER CONSUMPTION - PALC nodesS20S30L20L30XL20XL30Seq20Seq30 51121.826.244.228.541.627.445.2 105.611.215.132.327.443.317.236 153.87.510.322.323.941.611.725.8 202.85.77.819.6 38.28.919.7 Power Consumption – PALC vs. Round robin Experiments ran on the local IaaS cloud architecture 32
33
CURRENT RESEARCH POWER CONSUMPTION – PASC - RELATED Hasebe, et al. 2010 Power-Savings in large scale storage systems Prada, et al. 2012 Power-Aware Storage Architecture for HPC We provide macro management of storage devices based on current user utilization. Power down storage nodes during low utilization 33
34
CURRENT RESEARCH POWER CONSUMPTION - PASC PASC Persistent storage is necessary in cloud environments VMs are inherently stateless. How can we decrease the power consumption of storing persistent data without affecting availability? 34
35
Algorithm PASC Active: if user j becomes active: locate data j on cold storage node HS a current utilization of hot storage node a if HS a can accommodate data j +(Quota j – data j ): power on cold storage node transfer data j hot storage node a delete data j from cold storage node power off cold storage node else: //no hot storage node can accommodate data j perform cold-hot conversion In-Active: if user j becomes in-active: CS a current utilization of cold storage node a if CS a can accommodate data j : power on cold storage node a transfer data j cold storage node a delete data j from hot storage node power off cold storage node a Cold-Hot Conversion: CS a current utilization of cold storage node a for CS a ε [m] do: minUtil (if CS a utilization < minUtil) update CS a with minUtil to hot storage node perform In-active on CS a to transfer unused data to cold storage nodes Hot-Cold Conversion: for CS a ε [m] do: if ((CS a + CS a+1 utilization) < Threshold): transfer data CS a CS a+1 delete data from CS a update CS a to cold storage node Backup: if day < 7: incremental backup else: full backup 35
36
CURRENT RESEARCH POWER CONSUMPTION - PASC Typical Job Schedule (# of VM requests) 36
37
CURRENT RESEARCH POWER CONSUMPTION - PASC Experiment setup: 2 job schedules Typical Random 3 job types Web Server Network Intensive Virtual Cluster CPU Intensive Database Server Disk I/O Intensive 37
38
CURRENT RESEARCH POWER CONSUMPTION – PASC Efficiency: PASC vs. Always OnEfficiency: PASC vs. Always on (Random Schedule) 38
39
CURRENT RESEARCH POWER CONSUMPTION – PASC Power Consumed: PASC vs. Always On Power Consumed: PASC vs. Always on (Random Schedule) 39
40
CURRENT RESEARCH POWER CONSUMPTION - PASC Number of VMs WebClusterDatabase 10 13.22218.4 20 14.824.821.4 30 14.323.919.3 60 21.335.529.8 120 36.560.951.7 Power Consumption – PASC vs. Always on 40
41
CURRENT RESEARCH VIRTUAL MACHINE LIVE MIGRATION - RELATED Wu et al. Performance Models of VM live migration Li et al. VM live migration based on performance predictions Kuno et al. VM performance during migration 41
42
CURRENT RESEARCH VIRTUAL MACHINE LIVE MIGRATION Live migrations allow vm’s to be relocated Performance/Issues with migrating virtual machines: How does the vm perform during migration? How well does the cloud architecture perform the vm migration? How much power is consumed during migration? Live Virtual Machine Migration Three phases Push phase Pre-copy termination phase Pull-and-terminate phase 42
43
CURRENT RESEARCH VIRTUAL MACHINE LIVE MIGRATION 43
44
CURRENT RESEARCH VIRTUAL MACHINE LIVE MIGRATION Experiments covering the following three areas: Performance of Cloud Resources during Live Migration Performance of VM during Live Migration Power Consumption of VM Live Migration Setup: 3 VM types: Web: Apache Web server with PHP Network Intensive Cluster: Compute Pi to N th place CPU Intensive DB: MySQL Database Disk I/O Intensive 3 VM sizes 1. 256MB RAM, 2GB Disk, 1CPU core 2. 512MB RAM, 5GB Disk, 1 CPU core 3. 1024MB RAM, 10GB Disk, 2 CPU cores 44
45
CURRENT RESEARCH VIRTUAL MACHINE LIVE MIGRATION – CLOUD PERFORMANCE 45
46
CURRENT RESEARCH VIRTUAL MACHINE LIVE MIGRATION – VM PERFORMANCE 46
47
CURRENT RESEARCH VIRTUAL MACHINE LIVE MIGRATION – POWER CONSUMPTION 47
48
CONCLUSION Local cloud architectures are a viable alternative for organizations. Hosting a local cloud architecture can be expensive Need to reduce costs of owning/operating. Need to make resources easily available. Geographically distributed IaaS architecture. Low public networking resources Web based client/administrative interface. PALC PASC Reduce cost by actively migrating virtual machines 48
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.