Joint work: Yuezhi Zhou, Yaoxue Zhang, Tsinghua University, China Yinglian Xie, Carnegie Mellon University Virtual Disk based Centralized Management for.

Joint work:Yuezhi Zhou, Yaoxue Zhang, Tsinghua Uni

versity, ChinaYinglian Xie, Carnegie Mellon University

Virtual Disk based Centralized Management for Enterprise Networks

Challenge of Enterprise Systems

• Management of enterprise network systems based on PCs is still a big challenge:– PC: Full function desktop computer with

native software and data

– Software maintenance

– SecurityBecome more complicated with diverse types of OSes and applications coexisting !

Educational Classrooms

1 Every machine has to be installed with OSes and applications2 System states must be cleaned for each next class

Military Environments

1 Devices can access software and data only in a limited physical area2 No software or data can be carried outside the network boundary

Failed!!!Failed!!!

Existing Management Tools

• Automatically pushing installation images and patches– Examples: Marimba, Ghost

• Difficult to maintain consistency across machines– Vulnerable to errors or attacks due to the existence of

local data– Out of centralized control

Failed

Why Centralized Management

• Distributed diskless thick clients, yet centralized repositories of all software and data– Reduced software maintenance time– Enhanced security– Availability – Heterogeneous OS and application support– Easy software migration– Easy data backup and recovery

Why Not Thin Clients

• Performance– Poor scalability due to centralized

computing– Not appropriate for CPU/memory

intensive applications

• Cost– Need powerful server– Can not leverage the cheap and powerful

computing resources of clients

TransCom System Overview

Delivery network

TransCom Clients

…

…

TransCom Clients

Software repositories Data repositories

Servers Servers

Client: bare-hardware like computing platformSever: Regular desktop computersConnected by Ethernet

One such server can support 30-50 clients

Virtual Disk Concept

• Simulate traditional disks, with disk images holding the actual contents on the server

• Support heterogeneous OSes and applications transparently

Boot, Sharing, and Isolation

• Remote OS boot– Launch BIOS-enabled Vdisk access funct

ion first (replace INT 13H)

– Load OS, as if with regular hard disks

• Vdisk sharing, isolation, and recovery– Use different types of Vdisk

– Copy On Write (COW) for system image protection and recovery

Implementation and Deployment

Before AfterMaintenance time 4-8 hours per

week30 minutes per week

Availability 4-8 hours service down time every Thursday

No service interruption

Security Virus found, physical theft

No virus and worms found, no physical theft

Location: Central South University of Forestry & Technology in ChinaUsage: e-learning classroom for online EnglishDuration time: from May 2005 to July 2006 Numbers of clients: 30

Implementation: prototype system supporting Windows & Linux

Testbed Performance

• Compared with regular PC– One client case is better – Ten clients case is comparable

• Compared with thin-client systems (e.g., Citrix, RDP, and VNC)– Application performance (slow-motion, one client)

• Web browsing: reduce access latency 2-3 times• Video playback quality: improve 2-20 times

– Scalability (i-bench, synchronously)• Achieve almost constant latency as opposed to the thin-client

systems where latency grows linearly

Related Work

• Network computers– Proposed by Oracle, Sun, IBM, Apple, etc.– Can not support commodity OS and applications

• Thin-client systems– Sun Ray 1 [Sun Micro], RDP [Microsoft], ICA [Citrix] – Centralized computing and storage, need high-end servers

• Networked file systems– NFS [Sandberg, 1985] & AFS [Howard, 1988]– Can share user data; hard to share heterogeneous OSes

• Virtual machine based approaches– Collective [Chandra, 2005], ISR [Kozuch, 2004] and SoulPad [Cac

eres, 2005]– Can not achieve native performance

Summary and Future Work

• Centralizing both software and data reduces the management complexity of enterprise networks

• An example prototype: TransCom– Reduce maintenance time and effort– Achieve similar performance to PCs

• Future work– Support more types of OSes and devices– Optimize performance

Thanks!