s VLDB arena

Document Sample
s VLDB arena Powered By Docstoc
					   AReNA: Adaptive Distributed Catalog
    Infrastructure Based On Relevance
                  Networks

Vladimir Zadorozhny, University of Pittsburgh, Pittsburgh, PA
Avigdor Gal, Technion, Haifa
Louiqa Raschid, University of Maryland, College Park, MD
Quiang Ye, University of Pittsburgh, Pittsburgh, PA




                Nebula Project: http://db.sis.pitt.edu/projects/Nebula

                                                                         1
          Networked Query Processing

                     query

                                                         Relevant statistics: response time,
                                                         network delay, data transfer rate, etc.
                                      optimization
output                                                                       Network is not
                                                                            (well) predictable
                                                            Statistics        Statistics is not
                    evaluation                              about data
                                                                                  reliable




                         data




   Data sources are remote, distributed, heterogeneous
Networked Queries with Distributed Catalog

            query



                        optimization      Statistics
  output                                  about data


           evaluation




              data

                                       Scalability ?
Profile-Based Performance Monitoring




                             LEGEND:
                                 performance
                                 monitor
                                 content server

                                 client

  PM Aggregation ?               performance
                                 profile-based
                                 cluster
              Aggregated Latency Profiles
  A client/server pair is characterized by Individual Latency
  Profiles (iLP). iLPs capture latency distributions experienced by
  clients when connecting to a server.
                                                               12   23 
                   10 20                             iLP2 = 
           iLP1 =                                             0.35 0.65
                  0.3 0.7
                          
                                                                       



                                45 90 120
                      iLP3 =   0.2 0.3 0.5 
                                           


                                                              0.8
Similar non-randomly associated                 iLP1                         iLP2
iLPs are aggregated in Relevance
Networks
                                                       0.2                  0.2

iLP similarity measures:                                                          6
                                                                 iLP3
Correlation and Mutual Information
       Discovering Non-random Associations
          with Relevance Networks (RNs)
                                                Threshold=0.4
We adopt RNs as a                 0.75
management tool, to         LP1          LP3
manage large
numbers of iLPs.
                      0.9         0.5      0.45

                            LP2           LP4
                                   0.8
                                                Threshold=0.7

                                  0.75
                            LP1          LP3



                      0.9

                            LP2           LP4           7
                                   0.8
Relevance Networks




                     8
                   AReNA: Architecture
AReNA dynamically analyzes and visualizes meaningful relationships among
client/ server pairs using Relevance Networks (RNs). Relationships are
evaluated using passive measurements made by client applications and
gathered on a continuous basis. RNs allow AReNA managing thousands of
constantly changing iLPs

         V          Performance Prediction
         I
         Z
         U       RN Generation and Analysis
         A
         L
         I              Data Preparation
         Z
         E
         R               Data Collection

                                                   Large-Scale Experimental Testbeds
                                                         CNRI Handle System
                                                         PlanetLab Overlay
                                                   Around 50 000 Latency Profiles
                                                                              9
AReNA: Screenshot




                    10
        Demo
Tuesday: 16:00-17:30
  Friday: 09:00-10:30




                        11

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:0
posted:4/6/2013
language:English
pages:10
dominic.cecilia dominic.cecilia http://
About