Sho Niboshi
       Hitoshi Oi
   University of Aizu
{nibo,hitoshi}@oslab.biz
 Research Objective
 Conventional Approach
 Idea of Resource Control
 Methodology
 Experiment
 Results
 Conclusion
...
 Proposes effective resource
  management method in a virtualized
  system under dynamic workload
 What is a good resour...
 Conventional                 method allocates
 resources                   based on each VM’s
 demand:
   Performance


...
 Allocate       resources based on
               relative performance
                Resource                          ...
Mail Domain                       App Domain

      Mail server                      App server

    ① DT          LR     ...
 QoS   controller
   Purpose: determine the amount of
    resources to maintain acceptable
    performance
   Method: u...
 Arbitration     Controller
   Purposes : balances relative
    performance of domains
   Method:  adjusts to system ca...
    Resource separation takes some overheads
                                           (unused resources)
Resource Separ...
Domain            QoS indicator          Workload
Mail domain       Delivery Time          SPECmail2001
                  ...
 Control CPU time
 Experiments with two workload
  scenarios
 Compare
   WRC (With Proposed Resource Manager)
 Interva...
Default   WRC                                       Default   WRC
                      50                                ...
Default   WRC                                       Default   WRC
                      6                                 ...
Mail
                                                                                                                     ...
Mail
                                                                                                                     ...
•   Proposed effective resource manager
    –   Resource allocation reflecting application
        states
    –   Better p...
17
Upcoming SlideShare
Loading in...5
×

Xs sho niboshi

938

Published on

Published in: Technology
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
938
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
15
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Transcript of "Xs sho niboshi"

  1. 1. Sho Niboshi Hitoshi Oi University of Aizu {nibo,hitoshi}@oslab.biz
  2. 2.  Research Objective  Conventional Approach  Idea of Resource Control  Methodology  Experiment  Results  Conclusion 2
  3. 3.  Proposes effective resource management method in a virtualized system under dynamic workload  What is a good resource manager?  Maintains application performance with minimal resources  Balances performance between applications 3
  4. 4.  Conventional method allocates resources based on each VM’s demand: Performance Performance Resource Allocation Resource Allocation 4
  5. 5.  Allocate resources based on relative performance Resource Resource Performance Performance Mail Domain App Domain Mail Domain App Domain Keeps the performance balanced! 5
  6. 6. Mail Domain App Domain Mail server App server ① DT LR TT ② QoS Controller QoS Controller ③ Req Req Arbitration ④ Controller Control Domain ⑤ Allocates CPU time 6
  7. 7.  QoS controller  Purpose: determine the amount of resources to maintain acceptable performance  Method: utilizes the empirical relationship between the performance and the allocated resources  modeled by Fuzzy control theory which incorporates the experience of a human process in its control design 7
  8. 8.  Arbitration Controller  Purposes : balances relative performance of domains  Method: adjusts to system capacity and builds resource capacity layout Example: Domain A, Domain B Request: 60% + 90% = 150% Allocation:40% + 60% = 100% 8
  9. 9.  Resource separation takes some overheads (unused resources) Resource Separation  Select most efficient capacity layout Cap-all Q1 ≠ Q2 Reqtotal < 100% Q1 ≠ Q2 Reqtotal < 100% Cap-some Cap-none Q1 ≠ Q2 ※Q1=QoS in domain1 Resource Utilization Q2=QoS in domain2 Xen Summit AMD 2010 9
  10. 10. Domain QoS indicator Workload Mail domain Delivery Time SPECmail2001 Login Rate Java app domain Transaction Time SPECjbb2005 System Specification CPU AMD Athlon x2 (Dual Core) 2.0 GHz Main Memory 4GB HDD SATA 7200rpm x2 OS CentOS 5.2 (kernel 2.6.18) Virtualization Xen hypervisor 3.2.2 Software 10
  11. 11.  Control CPU time  Experiments with two workload scenarios  Compare  WRC (With Proposed Resource Manager) Intervals Scenario ↑Mail Scheduler in Xen)  Default (Default Scenario ↑Java Sec Mail Java Mail Java 1-399 Medium Medium Medium Medium 400-1199 High Medium Medium High 1200-1600 Medium Medium Medium Medium 11
  12. 12. Default WRC Default WRC 50 0.25 Transaction Time (msec) Delivery Time (sec) 40 0.2 30 0.15 20 0.1 10 0.05 0 0 Mail increase Mail increase 12
  13. 13. Default WRC Default WRC 6 0.25 Transaction Time (msec) Delivery Time (sec) 5 0.2 4 0.15 3 0.1 2 1 0.05 0 0 Java increase Java increase 13
  14. 14. Mail App server server QoS Controller QoS Controller Mail - QoS vs. Request (↑Mail) Arbitration Controller DeliveryTime Request 60 80 % of total CPU (Request) Delivery Time (sec) 70 50 60  Resource Request 40 30 50 40 follows QoS states 20 30 20  Adaptation delay 10 10 occurs occasionally 0 0 0 200 400 600 800 1000 1200 1400 1600 Java - QoS vs. Request (↑Mail)  In low priority TransactionTime Request domain, some requests 0.3 60 don’t take account of TransactionTime (msec) % of total CPU (Request) 0.25 50 0.2 40 QoS state 0.15 30  Request is based on 0.1 20 current CPU consumption 0.05 10 0 0 0 200 400 600 800 1000 1200 1400 1600 Xen Summit AMD 2010 14
  15. 15. Mail App server server QoS Controller QoS Controller Arbitration Controller Consumption / Request (↑Mail) Ratio (Mail/Java) 1.2 1.1 ratio 1  The ratio between 0.9 Consumption and 0.8 0 500 1000 1500 Request is quite Consumption / Request Distribution unstable 1.2 Consumption / Request (Java) 1.1  Different capacity 1 layouts lead to both Caps are 100 0.9 both caps are not 100 different resource only Mail's Cap is 100 distributions 0.8 onlyJava's Cap is 100 0.7 0.7 0.8 0.9 1 1.1 1.2 Consumption / Request (Mail) Xen Summit AMD 2010 15
  16. 16. • Proposed effective resource manager – Resource allocation reflecting application states – Better performance than conventional method in one experiment • However, worse performance in the other experiment • Future work – Fast adaptation – Uses variation of QoS state as control parameter – Dynamic rule construction – Fair resource allocation – Needs more consideration about cap=100 (=0) 16
  17. 17. 17
  1. A particular slide catching your eye?

    Clipping is a handy way to collect important slides you want to go back to later.

×