Scribd Logo
Upload

Welcome to Scribd!

  • 37 views

    Vmeter: Power Modelling For Virtualized Clouds

    Uploaded by

    rana

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Vmeter: Power Modelling For Virtualized Clouds

    Uploaded by

    rana
    37 views18 pages

    Original Title

    166779be97d6e92e2d98f2bb6b4ee8c20dc6

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    37 views18 pages

    Vmeter: Power Modelling For Virtualized Clouds

    Uploaded by

    rana

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 18

    VMeter: Power Modelling for Virtualized Clouds

    Ata E Husain Bohra and Vipin Chaudhary

    Department of Computer Science and Engineering,


    University at Buffalo, State University of New York, New York, U.S.A.

    International Parallel & Distributed Processing Symposium (IPDPS) © 2010 IEEE


    Outline
    • Introduction
    • System Events and Current Limitations
    • Power Model
    • Experimental Setup
    • Conclusion and Future Work

    2
    Introduction (1/2)
    • Most of the recent power monitoring techniques:
    – Use an externally attached power meter to obtain the
    power consumption for the complete system.
    – Lack a fine-grained mechanism to profile the power of
    an individual hosted VM.
    • Seamless job/VM migration can reduce the total
    power consumption:
    – Without a fine-grained data monitoring and analysis
    system, it may not guarantee.
    • VMeter is proposed to solve these problems.

    3
    Introduction (2/2)
    • The major contributions of this paper include:
    – Propose a low-overhead power model, VMeter, to
    predict instantaneous power consumption by individual
    VM.
    – Reduce the datacenter bill without attaching an external
    power meter.
    – Implement a utility using proposed power model on a
    cluster node.

    4
    System Events and Current Limitations
    (1/2)
    • Most hardware architectures expose system events via
    performance counter registers.
    – The access to these registers enables profiling of the
    complete system or the running applications.
    • We selected these events that have high correction with
    the power consumption to monitor vital system:
    – CPU_CLK_UNHALTED
    – DRAM_ACCESSES
    – INSTRUCTION_CACHE_FETCHES
    – DATA_CACHE_FETCHES
    – Plus use a disk monitor utility to sample disk accesses per VM
    • Because of the lack of any direct hardware performance counter to measure.

    5
    System Events and Current Limitations
    (2/2)
    • Limitations:
    – Monitoring tools cannot profile virtual disk reads when VM
    is mounted using blktap (Block Tap) drivers.
    – Furthermore, the tool is not capable of providing vital
    statistics for the cache and DRAM utilization per VM.
    – Virtualization package lacks a port of performance counter
    driver that can enable application level access to hardware
    performance counter registers.
    • Solution:
    – Develop a utility program to provide utilization details: CPU,
    cache, DRAM and disk for each VM.
    – Monitor by accessing system event files and hardware
    performance counters.

    6
    Power Model (1/5)
    • VMeter consists of two sub components:
    – vm_monitor
    – vm_power

    • VM_Monitor
    – Oprofile: To sample the hardware performance counters
    – iostat: To sample the disk accesses
    1. Run the above profilers periodically
    2. Parse or analyze the event data we want
    3. Generate the resource utilization per VM about CPU, cache,
    DRAM and disk

    7
    Power Model (2/5)
    • VM_Power
    – Because there are no means to measure actual VM
    power, we developed several models to predict it.
    • Objectives:
    – Low runtime overhead cost
    – Acceptable average deviation from actual measured
    power using external attached power meter

    8
    Power Model (3/5)
    • The more the utilization of the sub-components, the
    more is the total power consumption.
    • We created a simple 4-D linear weighted power model:
    𝑃 𝑡𝑜𝑡𝑎𝑙 = 𝑐0 + 𝑐1 𝑝𝐶𝑃𝑈 + 𝑐2 𝑝𝑐𝑎𝑐ℎ𝑒 + 𝑐3 𝑝𝐷𝑅𝐴𝑀 + 𝑐4 𝑝𝑑𝑖𝑠𝑘
    – The mean prediction accuracy is 82%
    – Can’t determine the power consumption by individual VM
    • We performed principal component analysis (PCA) on
    the input data set.
    – Result shows a high correlation between the pairs {CPU,
    cache} and {disk, DRAM}.

    9
    Power Model (4/5)
    • We can classify the system workloads into two
    categories: CPU-bound and I/O-bound process.
    • We then decomposed the linear model into two
    linear regression models:
    𝑃*𝐶𝑃𝑈,𝑐𝑎𝑐ℎ𝑒+ = 𝑎1 + 𝑎2 𝑝𝐶𝑃𝑈 + 𝑎3 𝑝𝑐𝑎𝑐ℎ𝑒
    𝑃*𝐷𝑅𝐴𝑀,𝑑𝑖𝑠𝑘+ = 𝑎4 + 𝑎5 𝑝𝐷𝑅𝐴𝑀 + 𝑎6 𝑝𝑑𝑖𝑠𝑘
    𝑃(𝑡𝑜𝑡𝑎𝑙) = α𝑃*𝐶𝑃𝑈,𝑐𝑎𝑐ℎ𝑒+ + β𝑃*𝐷𝑅𝐴𝑀,𝑑𝑖𝑠𝑘+
    – This model is better than the simple linear model.

    10
    Power Model (5/5)
    • We compute the power consumption for each VM,
    and decompose the total system power consumption
    into two categories:
    – The baseline and the dynamic power.
    𝑃*𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒+ = 𝛼 ∗ 𝑎1 + 𝛽 ∗ 𝑎4
    𝑁
    𝑃(𝑡𝑜𝑡𝑎𝑙) = 𝑃*𝑏𝑎𝑠𝑒𝑙𝑖𝑛𝑒+ + 𝑃*𝑑𝑜𝑚𝑎𝑖𝑛(𝑘)+
    𝑘=1
    𝑃*𝑑𝑜𝑚𝑎𝑖𝑛(𝑖)+ = 𝛼 𝑎2 𝑝𝐶𝑃𝑈(𝑖) + 𝑎3 𝑝𝑐𝑎𝑐ℎ𝑒(𝑖) + 𝛽 𝑎5 𝑝𝐷𝑅𝐴𝑀(𝑖) + 𝑎6 𝑝𝑑𝑖𝑠𝑘(𝑖)

    – This model fits each benchmark better than all earlier


    models.
    • Still a linear degree (low cost)

    11
    Experimental Setup (1/6)
    • Hardware:
    – AMD Opteron Sun Sparc Server Platforms
    • CentOS 5.3 (Kernel 2.6.18)
    • Xen-3.1 package
    – oprofile-0.9.3 + xenoprofile

    – WattsUp Pro power meter

    • Benchmark software:
    Name CPU Intensive I/O Intensive
    NAS-NPB 2.4 ● ●
    Iozone ●
    Bonnie++ ●
    BYTEmark ●
    Cachebench ● ●
    Dense matrix ●
    multiplication
    GCC ● ● 12
    Experimental Setup (2/6)

    4-D linear power model is used to predict instantaneous full system power. 13
    Experimental Setup (3/6)

    Our proposed power model is used to predict instantaneous full system power. 14
    Experimental Setup (4/6)
    I/O CPU

    CPU I/O

    CPU CPU

    CPU-intensive job consumes more power. 15


    Experimental Setup (5/6)

    I/O+CPU CPU+CPU

    I/O+CPU
    CPU+CPU
    I/O+I/O

    I/O+I/O

    {I/O + CPU}, {CPU + CPU}, {I/O + I/O} Comparison 16


    Experimental Setup (6/6)

    Housekeeping activity

    Increase the overhead by ~0.0121% 17


    Conclusion and Future Work
    • VMeter can:
    – Predict instantaneous full system power consumption
    for virtualized nodes with an average mean and median
    accuracy of 93% and 94%.
    – Generate power consumption profiles for all active
    domains on physical nodes, adding a very low 0.0121%
    overhead.
    • Next:
    – Extend the power model to include network sub-
    component.
    – Modify a job scheduler to implement the resulting
    power-saving and QoS ensuring policies.

    18

    You might also like