An Automated Profiling Subsystem for QoSAware Service

1
An Automated Profiling Subsystem for QoS-Aware
Service

• Tarek F.Abdelzaher
• Department of Computer Science
• University of Virginia

2
Outline

• Introduction
• Model and Architecture
• Implementation
• Application
• Conclusions and Future Work

3
Introduction

• Traditionally,real-time computing deal with
  special-purpose software running on closed
  embedded platforms .
• Today,the need for predictable QoS-sensitive
  computing is arising in a myriad of new
  general-purpose applications that run on a vast
  variety of commercial platforms.
• In this paper,it suggest an approach for
  developing real-time systems,in which the system
  determines the execution requirements of its
  tasks automatically via self-profiling.

4
Model and Architecture

• The Execution Model
• The goal of the automated profiling subsystem is
  to estimate the resource requirements of the
  service.
• Task i requires Ci j units of resource j during
  
• Its execution.If task Ti sends xi bytes of
  data ,its requirements for resource j are given
  by
• A data-size independent part
• B date-size dependent part
• ei modeling error

5
Model and Architecture(cont.)

• We can show from Eq.1 that the utilization Ucj
  of resource j due to traffic class c is
• Rc is the task arrival rate due to
  traffic of class c
• Wc is the total bandwidth of that
  class
• Eav is the mean modeling error
• The optimal estimates of Aj and Bj are those
  which reduce Eavj to zero

6
Model and Architecture(cont.)

• The total amount of resource j needed for task Ti
  of type k is denoted by Ci,kj
• The generalized server execution model equations
  become

7
Model and Architecture(cont.)

• Eq.(4) can be used to describe the necessary
  per-class resource allocation for admission
  control purpose.
• If consider only the total utilization ,task
  rate,and bandwidth aggregated over all traffic
  classes.
• Aggregating Eq(4) over all traffic classes ,we
  get

8
Model and Architecture(cont.)
9
Implementation

• The subsystem was implemented in C,and tested on
  a PC-based Linux platform.for the purpose of
  testing,an Apache 1.3.9 web server was used.

10
Application

• To investigate the performance of the profiling
  mechanism we profiled an Apache web server
  running on Linux.
• Two other machines were used to run clients that
  test the server with a synthetic workload.
• The workload is composed of a series of static
  files of lengths 1k through 64k.
• We used a web-load generator,called httperf 21
  on the client to bombard the server with web
  requests at different rates for files of
  different sizes.

11
Application cont.
12
Application cont.
13
Application cont.
14
Conclusions

• A simple service execution model was proposed,and
  a profiling library was designed and implemented
  that uses a least squares estimator to determine
  model parameters.
• This profiling information can be used in
  admission control and real-time scheduling
  decisions.
• Testing results qualitatively demonstrate the
  convergence of the estimation algorithm to stable
  parameter values.
• Inclusion of profiling support does not need
  architectural modifications to future web servers

15
Future Work

• One topic of interest is that of automatic
  identification of the execution model structure.
• in general,the number of parameters involved in
  the model might not be known.the model might also
  contain certain nonlinearities.
• Ensuring sufficient workload variability is an
  interesting issue.