Cache conscious index mechanism for main-memory databases
0 likes710 views
This document proposes the Cache-Sensitive T-tree (CST-tree), a cache-conscious index structure for main-memory databases. The CST-tree aims to improve cache behavior by decreasing node size to increase cache hits compared to traditional T-trees. It does not store the entire middle key array in each node to save space. Experimental results show the CST-tree provides better performance than T-trees in main memory databases, with performance between CSB+-trees and full CSB+-trees. It also uses less space than full CSB+-trees.
Cache conscious index mechanism for main-memory databases
- 1. Cache-Conscious Index Mechanism for Main-Memory Databases Sun Limei Email:[email protected] Shenyang Jianzhu University
- 2. Outline 1. Introduction 2. Related work 3. Cache sensitive T-trees 4. Experimental results 5. Conclusions
- 3. Motivation The speed gap between CPU and memory Improve cache behavior in main memory data processing
- 4. Cache Conscious Indexing CSS-tree(Cache Sensitive Search-Tree) CSB+-tree(Cache Sensitive B+-Tree) Segmented CSB+-tree Full CSB+-tree
- 5. Disadvantage It is considered that the size of CSB+-tree’s node should be equal to that of basic transferring unit There is a large discrepancy between the size of cache line and disk page Cache line size：32B-128B(MMDB) Disk page size：4K-64K(DRDB) CSS-tree and CSB+-tree both increase the height of B+-tree
- 6. About Cache Parameters: capacity, line(block) size and accociativity. Memory reference: hit or miss Line size: the basic transferring unit between cache and memory.
- 8. T-Tree’s Structure K’<K1<K2<….<Kn-1<Kn<K” parent ptr bf K1 K2 …...… Kn-1 Kn bf bf …...… K’ K” …...… Disadvantage: Only the minimum and the maximum keys are often used in comparison.
- 9. Cache Sensitive T-tree Decrease the node size and increase cache hit. No more space used because the middle key array is not needed all the time.
- 10. CST- Search Input：root，value Output：flag（if success flag=1，else flag=-1） Algorithm： 1. p=root; 2. while (p!=NULL) { 3. if (value<p’s minimum) p=p->LChild ; //search in p’s left subtree 4. else if(value>p’s maximum) p=p->RChild ; //search in p’s right subtree 5. else { //search inside p 6. if (value= =p’s minimum|| value= =p’s maximum) return 1; 7. else { 8. use binary search inside p’s middle-key array for key value; 9. if success return 1; 10. else return -1; 11. } 12. } 13. } 14. return -1; //empty tree
- 11. Search Time Analysis Let n denote the number of tuples being indexed, m denote the maximum number of keys inside a node, c denote the size of a cache line
- 14. Space Statistics
- 15. Conclusions CST-tree can provide better overall performance than T-tree in MMDB. CST-tree’s performance falls between CSB+- tree and full CSB+-tree. CST-tree saves more space than full-CSB+- tree.
- 16. Thank you!