Storing Trees on Disk Drives Medha Bhadkamkar, Fernando Farfan, - PowerPoint PPT Presentation

Storing Trees on Disk Drives Medha Bhadkamkar, Fernando Farfan, Vagelis Hristidis, Raju FLORIDA INTERNATIONAL UNIVERSITY Rangaswami 15th December, 2005 FAST 2005 WiP Report 1

Introduction  Tree data are becoming commonplace:  Offer an intuitive, natural way for organizing information.  Examples: XML, multi-res video, natural sciences data (e.g. Bioinformatics), even traditional directory-file hierarchies.  Disk drives are ubiquitous and seem irreplaceable  Current approaches:  Use relational databases  Use flat files  Our contributions  Examine the tree storage problem  Propose native data layout strategies for tree data 15th December, 2005 2 FAST 2005 WiP Report

Tree Structured Placement Idea: Optimize common accesses • Parent to child • Node to sibling Assumptions: • Each node occupies an entire disk block • Semi-sequential access information available 15th December, 2005 3 FAST 2005 WiP Report

Optimized Tree-Structured Placement  Problems with basic tree placement:  Significant fragmentation.  Large random seeks  Solution:  Use non-free tracks  Use rotationally-optimal track-regions 15th December, 2005 4 FAST 2005 WiP Report

Grouping  Sequential  Add nodes to ‘supernode’ until its capacity allows.  Use depth-first traversal to get next node  Low fragmentation  Tree-preserving  Groups adjacent nodes  Avoids cycles in original tree  Preserves original tree structure in grouping  Greater fragmentation 15th December, 2005 5 FAST 2005 WiP Report

Grouping Examples Sequential Tree-preserving Assumption: Supernode can fit 5 nodes 15th December, 2005 6 FAST 2005 WiP Report

Building Supernode Trees Sequential Supernode List • Uses sequential grouping • Nodes linked in the order they are created Tree-Preserving Supernode Tree • Uses tree-preserving grouping • Edges according to original tree Sequential Supernode Tree • Uses sequential grouping • Several possibilities for edge creation • Avoid cycles 15th December, 2005 7 FAST 2005 WiP Report

Performance Evaluation 15th December, 2005 8 FAST 2005 WiP Report

Future Work  Multiple drives  Modeling more complex data and access patterns  Allows data and application directed layout  Requires detailed model for the disk-drive  Storing graphs on disk drives…  More generic than trees!  Can use directed and weighted  Can model several data-types and access patterns  Can model relational data as well! 15th December, 2005 9 FAST 2005 WiP Report