1 / 30

CS6604 P ROJECT

CS6604 P ROJECT. E XAMINATION O F G RAPH P ARTITIONING S TORAGE METHODS F OR R OAD N ETWORKS Presenter: Andrew Connors Professor: Prof. Chang-Tien Lu. Abstract. Looking at networks with moving objects, in particular Road Networks

deiondre
Télécharger la présentation

CS6604 P ROJECT

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. CS6604 PROJECT EXAMINATION OF GRAPH PARTITIONING STORAGE METHODS FOR ROAD NETWORKS Presenter: Andrew Connors Professor: Prof. Chang-Tien Lu

  2. Abstract • Looking at networks with moving objects, in particular Road Networks • Use Graph Partitioning Scheme, but instead of basin on connectivity, base partitioning on road usage • Results in partitions that group nodes and edges based on most probable routes • Use these partitions to cluster data and use scheme similar to CCAM • Look at improvements in algorithms that track the user – as the system follows the user then nodes should be in loaded memory pages – reduce paging

  3. Approach • Take 2005 data from MnDOT who publish Average Annual Daily Traffic (AADT) for there entire network • Has 239166 nodes and 193096 edges • Use multilevel k-way partitioning schema • based on METIS schema of George Karypis • Adapt partitioning objective to be based on local probabilities • Use clustering based storage • Based on CCAM – but change from connectivity to probability • Instead of WCRR measure, use percentage page-misses – more direct measure

  4. Graph Partitioning • Is NP-Complete – so need heuristics • Multilevel partitioning involves these steps: • Graph coarsening phase – reduces number of nodes by collecting adjacent nodes and replace with aggregate node – use edge-matching – find maximal set of matching edges and collapsing vertices incident on each edge into one vertex • Initial partitioning phase – now this is done on a significantly smaller graph – use a recursive bisection – splitting in two balanced halves each time – this is where the real work is done – based on a heuristic and cost model • Un-coarsening phase – recover the original graph but keeping the partitioning intact

  5. Multilevel Partitioning

  6. Coarsening Phase • Coarsening is a well studied mechanism • Must be careful that it does not result in a sub-optimal partition • Solution is to ensure • No loss of information that is necessary for the partitioning algorithm • Noise is not added on un-coarsening • Do not reduce graph size too much – costs to much in performance and could result in too much information loss • Use edge-matching algorithm

  7. Edge Matching • Take an original Graph V0=V(V0,E0) with vertices V0 and edges E0 • On each step i, construct smaller a smaller graphs Gi=V(Vi,Ei) with vertices Vi and edges Ei where |Vi| > |Vi+1|

  8. Initial Graph Partitioning • Partitioning is NP-Complete and so need a good heuristic for performance • The Heuristic needs to achieve • Balancing and constraining the partition sizes • A cost functioning that when minimized (or maximized) results in the required partitions • Try to incorporate ratio-cut – which puts the partition sizes into cost – so can naturally achieve partitions with using constraints • Use recursive bisection – taking repeated bisections.

  9. Un-coarsening Phase • Run through back through graphs from coarsest to finest Gm-1, Gm-2,…..G0 • Assign nodes that were collapsed together in coarser graph to same partition in finer graph • However, may not have local minima at each step due to increase in degrees of freedom • Therefore, need refinement algorithm on each step

  10. Clustered Storage CCAM • Use same mechanism as in CCAM paper • Data clusters – in this case partitions are stored in same page

  11. Project Contribution • The important part is the graph partitioning scheme • Previous work: • minimizes the cost of the cut edges • leads to maximizing the Weighted Connectivity Residue Ratio (WCRR) • Indirectly, try to achieve the same goal of pages containing the most likely accessed data. • Derives pages of data where that data is related by being part of the most probable path of a object moving and constrained to a road network – i.e. uses the actual statistics directly

  12. Experiments - Design

  13. Experiments - Goal • Take 100 paths of 100 steps each from random starting points • Path following a walk allow high AADT edges to find most probable route • Use this to evaluate partitions by measure page miss ratio • Also look at paths formed from random works • Calculate WCRR and CRR from graph and partitions

  14. Measuring Performance • Do not need “actual” performance measures – like milliseconds etc. • Just need ratio of page hits to page misses - p • Introduce this metric into code to take measures directly • Related to WCRR:

  15. How WCRR and AADT Relate

  16. Experiments - Data • Took two subsets of data from MnDOT road network, focused on busy areas – i.e. near large city, different in order of magnitude to measure scalability • Used ESRIArcGIS software • MnDOT uses ESRI format • Students get free license from VT • However, not easy to export data in required format – took a long learning curve

  17. Complete MnDOT Data Only Minnesota 193,096 edges, 239,166 nodes

  18. Sub-sets MnDOT Data Took two subsets of data from MnDOT Medium 57,004 edges, 85,565 nodes

  19. Experiments - Data • Extracted to XML • Using ESRI GDBExchange Schema • Wrote Java code using Jakarta XMLBeans to import data and create Java objects • Wrote out files into a more usable and smaller CSV format • ESRI Network Graph exports do not retain edge-node mappings – only spatial references • Wrote more code to match edge start and end locations to node location • Produced node-edge mapping format • Used that to import into a Java based Graph storage format – adjacency list

  20. Experiments - Results

  21. Experiments - Results

  22. Experiments - Results

  23. Experiments – Random Walks

  24. Experiments – Random Walks

  25. Experiments – WCRR

  26. Experiments – WCRR

  27. Experiments – WCRR

  28. Experiments – WCRR

  29. Conclusions • Using AADT when generating partitions improved performance • But not for random walks • WCRR was an indicator of performance • Experimental and calculated values of WCRR correlated.

  30. Q + A Thank You!!

More Related