ehsan by huanghengdong

VIEWS: 2 PAGES: 9

									Survey of Datacenter Power
 Management Techniques
       Ehsan Nourbakhsh
      CS 7301 Spring 2010
 The University of Texas at Dallas
     Classification of Energy Costs
• Electricity Cost of Computation
   – Fan et al. [1]:      E   Pdt  t  ( P0  Pf  f )
                                                       3


• Electricity Cost of Idling t
   – Chen et al. [3]: may be up to 66% of the peak
     power
• Cost of CRAC Performance
   – Moore et al. [4]: eliminate 25% of the cooling
   – equal to $1 - $2 million annual cost reduction for a
     30; 000 sq ft datacenter.
                                                        2
         Thermal Management
• Moore et al. [4]
  – Zone-Based Discretization (ZBD)
     • Zones of neighbors have a power budget
  – Minimizing Heat Recirculation (MinHR)
     • Measure heat recirculation in datacenter
     • Need calibrations phases
• Heath et. al [5]
  – predict heat propagation
  – Mercury simulator, Feron and Feron-EC methods
                                                    3
  Component Power Management
• Dynamic Voltage Frequency Scaling (DVFS) [7]
• Felter et al. [8]*
  – Budgets per category (CPU, Memory, etc.)
• Elnozahy et al. [9]
  – independent voltage scaling (IVS)
  – Coordinated voltage scaling (CVS)
  – Vary on/vary off (VOVO)


                                                 4
       Virtual Machine Transfer
• Urgaonkar et al. [13]
  – application placement on a cluster of servers is
    NP-complete
• Moving the whole OS?
  – VM: Virtual Machine
  – PM: Physical Machine
• Khanna et al. [14]
  – Migrate VM with lowest utilization to PM with
    lowest available resources
                                                       6
VM Placement Based on Historical
            Data
• Bobro et al. [15]
  – Use historical data to predict trends
  – Use bin-packing to get most out of each server
• Chen et al. [16]
  – Queuing theory model for offered work-load        const

  – Control theory model for active feedback         Util. i
                                                     param
• Van et al. [17]
  – Formulate the problem as optimization problem
  – Reduce servers while honoring SLAs         Alloc. vector
                                                        7
VM Placement Based on Historical
          Data (2)
• Wood et al. [18]
  – Hotspot: the point that most workload is
    concentrated while resource are becoming
    insufficient




                                               8
 SoftStates and Global Awareness
• Nathuji et al. [19] (VirtualPower)
     • Use feedback from the operating system inside VM
     • Hypervisor: VM Manager (XEN[12], VMWare ESX [11])
     • PM-L: Local Policy, PM-G: Global Policy




                                                           9
    Network Bandwidth Concerns
• Meng et al. [22]
  – Traffic-aware VM Placement Problem (TVMPP)
  – correlation between the amount of traffic
    between machines in the same datacenter
  – suffering of other machines that are relying on
    those switches for communication
  – Method to assign VMs to PMs to reduce above
• Stage et al. [23]
  – Consider traffic resulting from VM transfers
                                                      10

								
To top