LiteGreen Saving Energy in Networked Desktops using Virtualization. Tathagata Das, Pradeep Padala, Venkat Padamanabhan, Ram Ramjee, Kang G.

Transcription

1 LiteGreen Saving Energy in Networked Desktops using Virtualization Tathagata Das, Pradeep Padala, Venkat Padamanabhan, Ram Ramjee, Kang G. Shin

2 Have you switched off your desktop?

3 Have you switched off your desktop? Key observation Users do not like disruption Automated energy saving is needed

4 Energy Usage of Electronics in the US (2006) [Greener PCs for the Enterprise, Nordman 06]

5 Energy Usage of Electronics in the US (2006) 250 TWh [Greener PCs for the Enterprise, Nordman 06]

6 Energy Usage of Electronics in the US (2006) 26% Enterprise PCs 250 TWh [Greener PCs for the Enterprise, Nordman 06]

7 Energy Usage of Electronics in the US (2006) 100 TWh 26% Enterprise PCs 14% Other PCs 250 TWh [Greener PCs for the Enterprise, Nordman 06]

8 Energy Usage of Electronics in the US (2006) 100 TWh 26% Enterprise PCs 14% Other PCs 250 TWh Other Electronics [Greener PCs for the Enterprise, Nordman 06]

9 How much energy can we save? Data gathering for ~120 users, MSRI, Summer 09 ~65000hrs worth of data

10 Total Idle Time (in minutes) How much energy can we save? Data gathering for ~120 users, MSRI, Summer 09 ~65000hrs worth of data Idle bins (in minutes)

11 Total Idle Time (in minutes) How much energy can we save? Data gathering for ~120 users, MSRI, Summer 09 ~65000hrs worth of data Many short idle periods (<60min) Idle bins (in minutes)

12 Total Idle Time (in minutes) Total Time (in Hours) How much energy can we save? Data gathering for ~120 users, MSRI, Summer 09 ~65000hrs worth of data <= Hrs 4200 > 3Hrs Idle bins (in minutes)

13 Total Idle Time (in minutes) Total Time (in Hours) How much energy can we save? Data gathering for ~120 users, MSRI, Summer 09 ~65000hrs worth of data Key observation Significant energy savings by exploiting short idle periods! <= Hrs 4200 > 3Hrs Idle bins (in minutes)

14 Existing Approaches and Related Work

15 Existing Approaches and Related Work Windows Power Management + Trivial inactivity easily handled Disruption to existing applications, cannot wakeup remotely Confused by spikes, night time traffic, background processes Protocol agnostic proxy [uses Wake-on-LAN mechanism] + Supports remote wakeup, stateless Manual, works only in subnet, disruption to existing applications Protocol proxy [Nedevschi et.al. NSDI 09] + Automatic: Protocol Specific Rules (ICMP, ARP, RDP, ) Disruption to existing applications, deployment problems Application proxy [Agarwal et.al. NSDI 09] + Reduces disruption to applications that have proxies Increased complexity in writing application specific proxies

16 Existing Approaches and Related Work Windows Power Management + Trivial inactivity easily handled Disruption to existing applications, cannot wakeup remotely Confused by spikes, night time traffic, background processes Protocol agnostic proxy [uses Wake-on-LAN mechanism] + Supports remote wakeup, stateless Manual, works only in subnet, disruption to existing applications Protocol proxy [Nedevschi et.al. NSDI 09] + Automatic: Protocol Specific Rules (ICMP, ARP, RDP, ) Disruption to existing applications, deployment problems Application proxy [Agarwal et.al. NSDI 09] + Reduces disruption to applications that have proxies Increased complexity in writing application specific proxies

17 Existing Approaches and Related Work Windows Power Management + Trivial inactivity easily handled Disruption to existing applications, cannot wakeup remotely Confused by spikes, night time traffic, background processes Protocol agnostic proxy [uses Wake-on-LAN mechanism] + Supports remote wakeup, stateless Manual, works only in subnet, disruption to existing applications Protocol proxy [Nedevschi et.al. NSDI 09] + Automatic: Protocol Specific Rules (ICMP, ARP, RDP, ) Disruption to existing applications, deployment problems Application proxy [Agarwal et.al. NSDI 09] + Reduces disruption to applications that have proxies Increased complexity in writing application specific proxies

18 Existing Approaches and Related Work Windows Power Management + Trivial inactivity easily handled Disruption to existing applications, cannot wakeup remotely Confused by spikes, night time traffic, background processes Protocol agnostic proxy [uses Wake-on-LAN mechanism] + Supports remote wakeup, stateless Manual, works only in subnet, disruption to existing applications Protocol proxy [Nedevschi et.al. NSDI 09] + Automatic: Protocol Specific Rules (ICMP, ARP, RDP, ) Disruption to existing applications, deployment problems Application proxy [Agarwal et.al. NSDI 09] + Reduces disruption to applications that have proxies Increased complexity in writing application specific proxies

19 Existing Approaches and Related Work Windows Power Management + Trivial inactivity easily handled Disruption to existing applications, cannot wakeup remotely Confused by spikes, night time traffic, background processes Protocol agnostic proxy [uses Wake-on-LAN mechanism] + Supports remote wakeup, stateless Key observation Manual, works only in subnet, disruption to existing applications Disruption to applications and lack of Protocol proxy [Nedevschi et.al. NSDI 09] generality are major limitations + Automatic: Protocol Specific Rules (ICMP, ARP, RDP, ) Disruption to existing applications, deployment problems Application proxy [Agarwal et.al. NSDI 09] + Reduces disruption to applications that have proxies Increased complexity in writing application specific proxies

20 Maintenance overhead Power management spectrum Energy savings

21 Maintenance overhead Power management spectrum Windows PM Automated Low savings Energy savings

22 Maintenance overhead Power management spectrum Windows PM Automated Low savings WoL Manual User disruption Energy savings

23 Maintenance overhead Power management spectrum Windows PM Automated Low savings Proxying Automated App stubs WoL Manual User disruption Energy savings

24 Maintenance overhead Power management spectrum Windows PM Automated Low savings Proxying Automated App stubs WoL Manual User disruption Energy savings Holy Grail

25 Maintenance overhead Power management spectrum Windows PM Automated Low savings Proxying Automated App stubs Key research question and goals How to automatically save energy? WoL Manual User disruption Minimal performance disruption Minimal maintenance Energy savings Holy Grail

26 LiteGreen: virtualization and consolidation

27 LiteGreen: virtualization and consolidation Hardware

28 LiteGreen: virtualization and consolidation Virtual Machine Virtualization Hardware

29 LiteGreen: virtualization and consolidation Virtual Machine Virtualization Hardware Server

30 LiteGreen: virtualization and consolidation Virtual Machine Virtualization Hardware Server Active VM Machine off Idle VM

31 LiteGreen: virtualization and consolidation Virtual Machine Virtualization Hardware Server Active VM Machine off Idle VM

32 LiteGreen: virtualization and consolidation Virtual Machine Virtualization Hardware Server Active VM Machine off Idle VM

33 LiteGreen Contributions

34 LiteGreen Contributions 1. Provide normal, undisrupted desktop experience How to make the machine usable instantly? How to mask the effect of VMs and their migration? Solution: Indirection (RDP) + live migration 2. Automated Energy Savings When to migrate? Which VMs to migrate? Solution: Savings algorithm driven by data analysis 3. Evaluation Hyper-V prototype and small-scale deployment Trace-driven analysis

35 LiteGreen Contributions 1. Provide normal, undisrupted desktop experience How to make the machine usable instantly? How to mask the effect of VMs and their migration? Solution: Indirection (RDP) + live migration 2. Automated Energy Savings When to migrate? Which VMs to migrate? Solution: Savings algorithm driven by data analysis 3. Evaluation Hyper-V prototype and small-scale deployment Trace-driven analysis

36 LiteGreen Contributions 1. Provide normal, undisrupted desktop experience How to make the machine usable instantly? How to mask the effect of VMs and their migration? Solution: Indirection (RDP) + live migration 2. Automated Energy Savings When to migrate? Which VMs to migrate? Solution: Savings algorithm driven by data analysis 3. Evaluation Hyper-V prototype and small-scale deployment Trace-driven analysis

37 LiteGreen Contributions 1. Provide normal, undisrupted desktop experience How to make the machine usable instantly? How to mask the effect of VMs and their migration? Solution: Indirection (RDP) + live migration 2. Automated Energy Savings When to migrate? Which VMs to migrate? Solution: Savings algorithm driven by data analysis 3. Evaluation Hyper-V prototype and small-scale deployment Trace-driven analysis

38 LiteGreen Contributions 1. Provide normal, undisrupted desktop experience How to make the machine usable instantly? How to mask the effect of VMs and their migration? Solution: Indirection (RDP) + live migration 2. Automated Energy Savings When to migrate? Which VMs to migrate? Solution: Savings algorithm driven by data analysis 3. Evaluation Hyper-V prototype and small-scale deployment Trace-driven analysis

39 Talk outline Background and Motivation What is LiteGreen? Design and Implementation (demo) Evaluation Trace-driven Analysis Hyper-V Deployment Limitations and Future Work Conclusion

40 System Architecture LiteGreen Server LiteGreen Controller Idle VM VMs Storage Server Gigabit Switch VM VM Idle Desktops Active Desktops

41 LiteGreen Operation

42 LiteGreen Operation Problem: How to mask the effect of migration?

43 LiteGreen Operation Problem: How to mask the effect of migration? Solution: RDP client + Live migration

44 LiteGreen Operation Solution: RDP client + Live migration Virtual Machine Virtualization Hardware Server Active VM Idle VM

45 LiteGreen Operation Solution: RDP client + Live migration RDP client Virtual Machine Virtualization Hardware Server Active VM Idle VM

46 LiteGreen Operation Solution: RDP client + Live migration RDP client Virtual Machine Virtualization Hardware Server Active VM Idle VM

47 LiteGreen Operation Solution: RDP client + Live migration RDP client Virtual Machine Virtualization Hardware Server Active VM Idle VM

48 LiteGreen Operation Solution: RDP client + Live migration RDP client Virtual Machine Virtualization Hardware Server Active VM Idle VM

49 LiteGreen Operation Solution: RDP client + Live migration RDP client Virtual Machine Virtualization Hardware Server Active VM Idle VM

50 LiteGreen Operation Solution: RDP client + Live migration Few seconds RDP client Virtual Machine Virtualization Hardware Server Active VM Idle VM

51 LiteGreen Operation RDP client Solution: RDP client + Live migration Virtual Machine Few seconds Live Migration Virtualization Hardware Server Active VM Idle VM

52 LiteGreen Operation Solution: RDP client + Live migration RDP client Virtual Machine Virtualization Hardware Server Active VM Idle VM

53 LiteGreen in action (demo)

54 Savings algorithm Intuition

55 Savings algorithm Intuition Active VM = (User is interactive) OR (Resource usage > ResThr)

56 Savings algorithm Intuition Active VM = (User is interactive) OR (Resource usage > ResThr) ACTIVE

57 Savings algorithm Intuition Active VM = (User is interactive) OR (Resource usage > ResThr) ACTIVE VM not active for IdleThr minutes MIGRATING

58 Savings algorithm Intuition Active VM = (User is interactive) OR (Resource usage > ResThr) VM is active migration is cancelled ACTIVE VM not active for IdleThr minutes MIGRATING

59 Savings algorithm Intuition Active VM = (User is interactive) OR (Resource usage > ResThr) VM is active migration is cancelled ACTIVE VM not active for IdleThr minutes MIGRATING migration finishes IDLE

60 Savings algorithm Intuition Active VM = (User is interactive) OR (Resource usage > ResThr) VM is active migration is cancelled ACTIVE VM not active for IdleThr minutes MIGRATING migration finishes IDLE VM is active migration back to desktop initiated

61 Savings algorithm Intuition Active VM = (User is interactive) OR (Resource usage > ResThr) VM is active migration is cancelled ACTIVE VM not active for IdleThr minutes MIGRATING migration finishes IDLE VM is active migration back to desktop initiated How to choose ResThr and IdleThr? Statistical analysis

62 Savings Algorithm - Policies idle VMs could be pulled into the server, subject to capacity [ eligible for pull list] active VMs must be pushed back to the desktop machine [ mandatory to push list] push pull

63 Savings Algorithm - Policies idle VMs could be pulled into the server, subject to capacity [ eligible for pull list] active VMs must be pushed back to the desktop machine [ mandatory to push list] push pull Two Policies 1. Conservative: Active VM = (User is interactive) OR (Resource usage > ResThr) 2. MaxSavings: Active VM = (User is interactive)

64 Talk outline Background and Motivation What is LiteGreen? Design and Implementation (demo) Evaluation Trace-driven Analysis Hyper-V Deployment Limitations and Future Work Conclusion

65 Trace driven analysis: Energy savings

66 Sleep Time (%) Trace driven analysis: Energy savings LG MaxSavings LG Conservative Existing

67 Sleep Time (%) Trace driven analysis: Energy savings LG MaxSavings LG Conservative Existing Data gathered at MSR India, 120 machines 65,000 user-hours of data Existing power management Manual switching off Windows PM plans Balanced (most users) 30 min IdleThr (unclear what idle means) Few users have high performance plan LiteGreen thresholds CPUThr = 10%, IdleThr = 5min

68 Sleep Time (%) Sleep Time (%) Trace driven analysis: Energy savings LG MaxSavings LG Conservative Existing Data Weekend gathered Dayat MSR India, machines WeekNight 5.72 WeekDay 65,000 user-hours 5.32 of data Existing power management Manual switching off Windows PM plans Balanced (most users) min IdleThr 4.66 (unclear what idle means) Few users have high performance plan LiteGreen 6.63 thresholds 0 Weekend Night CPUThr Existing = 10%, LiteGreen IdleThr = 5min Conservative LiteGreen MaxSavings

69 User who is always on

70 User who is always on Many short idle periods Spikes cause problems for existing PM

71 Sleep Time (%) User who is always on Weekend Night Weekend Day Week Night Week Day Existing LiteGreen

72 User who manually switches off Nothing can beat manual PM

73 User who manually switches off Sleep Time (%) Weekend Night Weekend Day Week Night Week Day Existing LiteGreen

74 Prototype with Hyper-V and Xen A 2-node server cluster One server node (32GB RAM + 8 cores CPU) running Windows 2008 R2 + Hyper-V R2 One storage node (2x100GB disks) accessed through iscsi 10 user desktops Mixture of researchers and non-researchers USENIX paper submission done using two of the desktops All nodes connected on 1Gbps LAN Memory ballooning and consolidation experiments with Xen

75 No. of Migrations Deployment results - # of migrations 50 Holiday Night Holiday Day Weekday Night WeekDay Day Days in Deployment

76 No. of Migrations Deployment results - # of migrations 50 Holiday Night Holiday Day Weekday Night WeekDay Day Days in Deployment LiteGreen can exploit idle periods during both day and night

77 No. of Migrations Deployment results - # of migrations Holiday Night Holiday Day Weekday Night WeekDay Day Total Energy Savings: 71% Days in Deployment LiteGreen can exploit idle periods during both day and night

78 Xen Memory Ballooning Effects 256MB Every 30 minutes memory of desktop VM reduced by 128MB Adaptive mechanism based on page faults can be developed

79 Talk outline Background and Motivation What is LiteGreen? Design and Implementation (demo) Evaluation Trace-driven Analysis Hyper-V Deployment Limitations and Future Work Conclusion

80 Limitations and Future work Virtualization GPU virtualization. Client hypervisors can help. Shared storage based live migration. Live migration with local disk state can help. Evaluating consolidation of idle VMs How much does memory sharing help? What metrics to be used for performance disruption? Combining with Software only approaches. SleepServer: A Software-Only Approach for Reducing the Energy Consumption of PCs within Enterprise Environments [Agarwal et.al. USENIX ATC 10] Proxying. Sleepless in Seattle No Longer [Reich et. al. USENIX ATC 10]

81 Summary: LiteGreen in one slide RDP client VM Virtualization Hardware Live Migration VM Server

82 Summary: LiteGreen in one slide RDP client VM Virtualization Hardware Live Migration VM Server 1. Significant energy savings exist in short idle periods

83 Summary: LiteGreen in one slide RDP client VM Virtualization Hardware Live Migration VM Server 1. Significant energy savings exist in short idle periods 2. Holygrail = High manageability + High savings

84 Summary: LiteGreen in one slide RDP client VM Virtualization Hardware Live Migration VM Server 1. Significant energy savings exist in short idle periods 2. Holygrail = High manageability + High savings 3. LiteGreen exploits live migration and RDP to gain short idle period savings

85 Summary: LiteGreen in one slide RDP client VM Virtualization Hardware Live Migration VM Server 1. Significant energy savings exist in short idle periods 2. Holygrail = High manageability + High savings 3. LiteGreen exploits live migration and RDP to gain short idle period savings 4. LiteGreen can save 72-74% more energy on top of existing savings (32%) for collected traces

86 Summary: LiteGreen in one slide RDP client VM Virtualization Hardware Live Migration VM Server 1. Significant energy savings exist in short idle periods 2. Holygrail = High manageability + High savings 3. LiteGreen exploits live migration and RDP to gain short idle period savings 4. LiteGreen can save 72-74% more energy on top of existing savings (32%) for collected traces 5. Hyper-V prototype shows total of 71% savings for 10 users

87 Summary: LiteGreen in one slide RDP client VM Virtualization Hardware Live Migration VM Server 1. Significant energy savings exist in short idle periods 2. Holygrail = High manageability + High savings 3. LiteGreen exploits live migration and RDP to gain short idle period savings 4. LiteGreen can save 72-74% more energy on top of existing savings (32%) for collected traces 5. Hyper-V prototype shows total of 71% savings for 10 users LiteGreen = Saving Desktop Energy using Virtualization