Download presentation
Presentation is loading. Please wait.
Published byAmberly McGee Modified over 8 years ago
1
1 Agility in Virtualized Utility Computing Hangwei Qian, Elliot Miller, Wei Zhang Michael Rabinovich, Craig E. Wills {EECS Department, Case Western Reserve University} {CS Department, Worcester Polytechnic Institute}
2
2 Internet Applications: Resource Provisioning Challenge Unpredictable demand Challenge: how much resource to provision? Too little – lose business Too much – poor utilization Utility Computing is a promising answer
3
3 Resource Management Alternatives Application server sharing Different applications running on the same app server OS-level sharing Each application runs in dedicated app server Several app servers share physical machine Dedicated physical machines Each physical machine runs one app server with one app Resources are assigned at the granularity of entire machines Isolation (security, fault, performance) vs. utilization tradeoff
4
4 An Appealing Option: Virtual Machines Almost as good isolation as physical machines reassignment Much better utilization than physical machine reassignment Heterogeneity support
5
5 Target Utility Computing Platform Applications run on dedicated app servers App servers run on dedicated VMs App instances form app-level clusters
6
6 Goal Improve the agility of Utility Computing “how quickly Utility Computing Platform can react to changing demands.”
7
7 VM-based resource management alternatives Migration of VMs Redeployment of Application Servers on Demand
8
8 Migration of VMs Start/Stop VMs Predeploy everywhere Start more instances in high-demand areas Stop underloaded instances Clone VM Stop overloaded VMs on overutilized hosts Copy stopped VMs to other underutilized hosts Start the cloned VM on target hosts Suspend/resume Pre-deploy VMs across Hosts Suspend VMs with Unpopular Applications Resume VMs with popular Applications Live VM Migration
9
9 Redeployment of Application Servers on Demand Pre-deploy VMs across Hosts Remove app servers with Unpopular Applications from VMs Add app servers with popular Applications to VMs
10
10 Our Approach: Ghost VMs Active VM: alive and utilized for application request processing Ghost VM: alive but unutilized Fully participate in cluster maintenance Do not process any requests Spend little CPU cycles
11
11 Resource Reassignment Mechanism Ghost implementation options: switch reconfiguration on demand -Remove VMs from switch -Set the Max Connection option to minimum (1) for the ghost VMs
12
12 Evaluation Experiment Environment Hardware: Intel (R) 4-core 2.33 GHz, CPU, 4G memory, 146G disk with 15K RPM Nortel Alteon 2208 Application Switch Software: Linux 2.6.17-10-generic SMP VMware Server 1.0.1 Websphere 6 Network Deployment
13
13 Migration of VMs Start/Stop VM About 64s to start a VM when its memory is flushed out, with a competing VM allocating memory moderately About 20s to start the cluster node agent About 97s to start the application server Totally about 181s Clone VMs About 15 minutes for a VM with 1G memory and 10G disk in a 100Mbps local network.
14
14 Migration of VMs (cont’ed) Suspend/Resume VMs About 5s on average in a repeated experiments About 10s on average for VM with memory swapped out About 14s on average with a competing VM (more realistic) About 30s for app server on the resumed VM to get re- integrated into the cluster Totally about 44s
15
15 Redeployment of Application Servers on Demand About 95s to stop a cluster member (application server) About 19s to create a cluster member About 97s to start a cluster member Totally about 211s
16
16 Agility of the Above Methods MethodAgility (sec) Migration of VMs~180 Redeployment of App Servers~210 Suspend/Resume VMs>44 The best agility we can expect is: 44s
17
17 Active/Ghost VMs Overhead of ghost VM: 1.4%-3.3% CPU Agility 7s: Add/remove VM from switch 2s: Set the Max connection to minimum(1)
18
18 Memory Swap-in Problem? For a VM with 1G memory about 652s to swap in all the memory about 267s to swap in 50% memory about 80s to swap in 5% memory Solution: Keep Ghosts in memory! Only run as many active+ghost VMs as total physical memory allows Also reserve memory for host OS (1G) Adding memory is much cheaper than adding physical machines Only very small number of ghost VMs are needed to hide resource reassignment latency VMware ESX: Reserved Memory for VMs VM state hierarchy Suspended Ghost Active
19
19 Evaluation Summary MethodAgility (sec) Migration of VMs~180 Redeployment of App Servers~210 Suspend/Resume VMs>44 Active/Ghost VMs2-7 Active/Ghost VMs has the best Agility The price is memory overhead Summary
20
20 Future Work Extend the scale of our test platform Move to VMware ESX More investigation of resource tradeoffs for ghost, suspended and active VMs Explore more on global and local data center resource management Build full-function prototype for resource management in Utility Computing
21
21 Summary Internet applications introduce resource provisioning challenge Utility computing is a promising paradigm VMs gain popularity as resource management approach Agility is an important characteristic of a utility computing platform Ghost VMs as a mechanism for agile utility computing
22
22 Thank you ! Questions?
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.