talk

Document Sample
talk Powered By Docstoc
					   Quantifying the Benefits of
    Resource Multiplexing in
   On-Demand Data Centers
Abhishek Chandra                             Pawan Goyal
 Prashant Shenoy                        IBM Almaden, San Jose
 UMASS Amherst




    UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science
    Motivation
   On-demand Data Centers
       Server farms
       Rent computing and storage
        resources to applications
       Revenue for meeting
        application workload levels

   Goals:
       Satisfy dynamically changing application requirements
       Maximize resource utilization of the platform
       Robustness against “Slashdot” effects


           UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   2
    Dynamic Resource Allocation
   Existing techniques:
       Oceano [Appleby01], HP Utility Data Center [Rolia00],
        MUSE [Chase01], COD [Doyle02], SHARC [Uragaon02]
       Differ in allocation policies and mechanisms
   Common features:
       Periodically re-allocate resources among applications
       Estimate workloads for near future
       Statistical multiplexing of resources

   Question: Which techniques work best and when?


           UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   3
On-demand Allocation: Practical Issues

   How often and how fine should the re-allocation
    be done?
   How well can the application requirements be
    estimated?
   How much “head room” should be allowed to
    absorb transient loads?
   Do large number of customers lead to better
    statistical multiplexing?




      UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   4
Talk Outline

   Motivation

   System Model and Metrics

   Performance Study

   Conclusions and Future Work




      UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   5
     System Model

   Cluster of servers
   Homogeneous pool of
    resources
   No constraints on
    application placement

   Time granularity (Δt): Period of re-allocation
       E.g.: re-allocate once every minute, hour, day
   Space granularity (Δs): Resource allocation unit
       E.g: re-allocate partial/whole server, server group


            UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   6
    Optimal Resource Allocation

   Infinitesimally small allocation granularity
   Allocates precise amount of resource
   No resource wastage



              Ropt

      Resource
      Allocation

                            Time


          UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   7
    Practical Resource Allocation
   Allocation done periodically and in fixed quanta
   Fixed resource allocation for next period
   Clairvoyant scheme: Predict peak application
    requirements for the next allocation period


                                 Δt
                         Δs

    Resource
    Allocation

                           Time


         UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   8
Capacity Overhead


        Rpract            ρ
          Ropt
Resource
Allocation

                       Time


                   Rpract  Ropt 
                                 100
                       Ropt      

     UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   9
Performance Study
    Workload:
      3 e-commerce traces

      24-hour long



     Workload        Number of        Avg. Request       Peak bit-rate
                     Requests             Size

    Ecommerce1       1,194,137           3.95 KB          458.1 KB/s


    Ecommerce2       1,674,672           3.85 KB          1631.0 KB/s


    Ecommerce3        251,352            7.24 KB          1346.9 KB/s




        UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   10
                              Effect of Allocation Granularity
                                  Time granularity                                                     Space granularity
                        140                                                                 200

                                  s=0.001                                                   175       t=1 min
Capacity Overhead (%)




                        120




                                                                    Capacity Overhead (%)
                                  s=0.01                                                              t=10 min
                                                                                            150
                        100       s=0.05                                                              t=1 hr
                                  s=0.2                                                     125       t=10 hrs
                        80
                                                                                            100
                        60
                                                                                            75
                        40
                                                                                            50
                        20
                                                                                            25
                         0                                                                   0
                              1   10    100     1000   10000 100000                          0.0001       0.001      0.01       0.1    1
                                   Time alloc unit (sec)                                                 Resource alloc unit (R_opt)

                   Fine time scale with reasonably fine resource unit desirable
                                   UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science                               11
Effect of Prediction Inaccuracy
                             140

                             120
     Capacity Overhead (%)




                             100

                              80

                              60

                              40                                                    t=1 min
                                                                                    t=10 min
                              20
                                                                                    t=1 hr
                                                                                    t=10 hrs
                               0
                                   0            50             100            150             200

                                              Prediction Inaccuracy (% std dev)

   Fine allocation is better even with inaccurate prediction

                             UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   12
Effect of Overprovisioning
                            300
                                      t=1 min
                                      t=10 min
    Capacity Overhead (%)



                            250
                                      t=1 hr
                            200       t=10 hrs

                            150


                            100


                            50


                             0
                                  0         20        40          60          80          100
                                                     Head Room (%)

   Finer allocation achieves same “head room” with less overhead


                             UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   13
Effect of Number of Customers
                   1000
                                    t=1 min
                        900
                                    t=10 min
Capacity Overhead (%)




                        800
                                    t=1 hr
                        700         t=10 hrs
                        600
                        500
                        400
                        300
                        200
                        100
                          0
                              0          20           40           60          80          100
                                                  Number of customers

                       Large number of customers provide more opportunity
                        for statistical multiplexing

                              UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   14
Data Center Architectures

   Dedicated
       Allocation of whole servers
       Typical reallocation in order of 30 minutes
   Shared
       Fractional server resources
       Reallocation in seconds or minutes
   Fast Reallocation
       Reserved server pools, remote booting
       Reallocation in a few minutes



        UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   15
    Comparison of Architectures
                                Optimal       Dedicated         Fast             Shared
Data Center     Number of        Reqmt       Architecture    Reallocation     Architecture
Configuration   customers       (Num of        (Num of        (Num of           (Num of
                                servers)       servers)       servers)         servers)


   Small             3             20             34               31                25


  Medium            15            100            388              304               148


   Large            30           1000            5017            3759               1739




            UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science            16
Implications and Opportunities

   Cost of re-allocation
       Partial server: ~1 syscall/min
       Full server: Rebooting, disk scrubbing, etc.
       Virtual machines: Low cost of reallocation with
        encapsulation
   Prediction:
       Work-conserving scheduler at fine time-scales
       Accurate prediction possible at minutes, hours




        UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   17
Conclusions and Future Work
   Dynamic Resource Allocation for data centers
   Fine allocation granularity desirable
       Even with inaccurate prediction
       To achieve more “head room”
   Large number of customers lead to higher
    multiplexing benefits

   Future Work:
       Effect of affinity, placement constraints
       Re-allocation overhead
       Stability of resource allocation

        UNIVERSITY OF MASSACHUSETTS, AMHERST – Department of Computer Science   18

				
DOCUMENT INFO
Shared By:
Categories:
Stats:
views:5
posted:3/5/2010
language:
pages:18