homework lab9 aras umit

Upload: aras-tarhan

Post on 06-Apr-2018

222 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/3/2019 Homework Lab9 Aras Umit

    1/17

    Fall 2011Homework Lab 9Faik Aras TarhanUmit Cavus Buyuksahin

  • 8/3/2019 Homework Lab9 Aras Umit

    2/17

    OVERVIEW

    To characterize Multi-tier enterprise software systems, it

    is important to:

    understand the peak capabilities of the system

    locate and quantify performance bottlenecks (such asCPU/disk/network)

    ensure the validity of the measured data

    These goals can be achieved by a systematic

    methodology which provides a basis for performance

    comparison of different e-business middleware

  • 8/3/2019 Homework Lab9 Aras Umit

    3/17

    Which Approach and Why ?

    Two approaches for systematic methodology

    Stochastic Analysis

    Queuing Operational Analysis

    Second one was chosen by the paper since;

    simple to measure for an N-tier system

    allows to make use of operational laws, which have been

    well understood for more than 30 years

    application of these fundamental equations can provide

    sanity- checks

  • 8/3/2019 Homework Lab9 Aras Umit

    4/17

    Queuing Model of a 3-Tier

    System

    consists of

    a web server ( a script (e.g. ASP or JSP) or Java Servlet runs )

    an application server (a business object (e.g. COM+, EJB))and a database server ( SQL,Oracle etc. )

    also a client to generate the workload (e.g. HTTP request)

  • 8/3/2019 Homework Lab9 Aras Umit

    5/17

    Key Metrics and Basic

    EquationsMetrics:

    the throughput in requests per second (X), a measure of

    work done by the system in terms of completed request

    the response time in seconds (R), the average latency of arequest

    the number of concurrent requests in the system (N)

    Equations:

    Utilization law: Ui = X * Di

    Littles Law:N = X * R

  • 8/3/2019 Homework Lab9 Aras Umit

    6/17

    Experiment Tools and Setup

    A stress test tool, called Web Application Stress tool(WAS), from Microsoft is used.

    composes the number of concurrent threads that willmake requests

    runs the script for some amount of time

    measures:

    performance data (e.g. CPU utilization, disk utilization )

    throughput for the test (X)

    (TTLB) response time for the test( If TTLB is big enough, client delay can be so negligible thatRand TTLB are considered as equal )

  • 8/3/2019 Homework Lab9 Aras Umit

    7/17

    General Calculation

    ProcedureService demands for CPU, Disk, Network by the

    Utilization Law

    The average of the measured service demands for eachresource is calculated

    The service demand values were inspected to find and

    evaluate the performance bottlenecks

    Littles Law is used to help sanity-check

  • 8/3/2019 Homework Lab9 Aras Umit

    8/17

    Experiment on Servlet A

    Cpu is the

    bottleneck

    The service

    demand is ~24 ms

    Derived X matches

    closely the

    measured

    throughput from

    WAS

  • 8/3/2019 Homework Lab9 Aras Umit

    9/17

    Experiment on

    Servlet AThe measurement is valid, as

    verified by the equations.

    The equations are applicable toreal world setups.

    We have a way of comparing

    workloads in terms of resource

    usage.

  • 8/3/2019 Homework Lab9 Aras Umit

    10/17

    Experiment on Servlet BThe deviation of the two curvesindicates that there is indeed a

    problem with the measurement

    Running a large number of

    threads on such a system meansthat the client system CPUs are

    spending too much time in

    context switching between

    threads, rather than generatingrequests to the server

    it is worthwhile to use Littles Law to validate measurements

  • 8/3/2019 Homework Lab9 Aras Umit

    11/17

    IIS Web Server Experiment (

    with 20 KB HTML File)

    the 100Mb/s network is the bottleneck resource

    Xmax= 1/1.54 = 647 req/sec which is within 5% of

    the measured peak throughput

  • 8/3/2019 Homework Lab9 Aras Umit

    12/17

    IIS Web Server Experiment (

    with 500 B HTML File)

    the CPU is the bottleneck

    Even with three clients, CPU does not load up to 90%

    Xmax= 1/Dmax while Dmax= 0.21ms Xmax= 4821req/sec

    The measured peak throughput is 4289 req/sec which is a

    discrepancy of about 11%

  • 8/3/2019 Homework Lab9 Aras Umit

    13/17

    Conclusion on IIS Web Server

    ExperimentIn second experiment the stress clients were unable tosaturate the web server

    Adding more client machines would reduce or eliminate

    this discrepancy or

    a larger file size would allow us to saturate the server,without adding more client systems

    To sum up;

    if we get a wider discrepancy between measurements andcalculated or expected data, we definitely need to dig deeperinto the causes for the discrepancy

  • 8/3/2019 Homework Lab9 Aras Umit

    14/17

    3-tier System Middleware

    Comparisonare commonly used to separate business logic from thepresentation layer at the front end, and enterprise dataaccess at the back end

    the CPU was the performance bottleneck for this test, thedata for the disk and network is not shown.

    Observe in the next slide that the overall throughput X is alot higher for Middleware II compared to Middleware I (352req/sec vs. 100 req/sec)

    It could be concluded that Middleware II is 3.5 times fasterthan Middleware I unless we dig deeper as concluded in theprevious slide

  • 8/3/2019 Homework Lab9 Aras Umit

    15/17

    3-tier System Middleware

    Comparison

  • 8/3/2019 Homework Lab9 Aras Umit

    16/17

    3-tier System Middleware

    Comparison ConclusionThe max throughput for Middleware is given by: Xmax

    = 1/Dmax= 1/9.6 = 104 req/sec so the measurements

    seem valid within a reasonable margin of error

    the service demand for the application server is nearly

    the same for both Middleware I and Middleware II

    both technologies are actually equivalent in their

    business logic performance!

    Middleware II has the advantage is in the performance

    of the web server compared to Middleware I

  • 8/3/2019 Homework Lab9 Aras Umit

    17/17

    Discussion

    Measurements must be validated against expectations byusing Littles Law

    If the measured and calculated values diverge significantly,further analysis is required

    The client delay which is also called think time for a 3-tiersystem affects the slope of the throughput curve

    As the delay is increased, the throughput curve shifts to theright

    The maximum throughput of the system is still the same:Xmax= 1 / Dmax