These results were obtained in a system with the characteristics: 50 GB TPC-H; 20
nodes, each with 866 MHz processor; 512 MB RAM). The alternatives compared
are: online??”every node is online; FPR??”fully partitioned replicas; PR(10)??”partitioned
replicas with two groups of 10 nodes; PR(5)??”partitioned replicas with 4
groups of 5 nodes. These results show the much larger penalty incurred by FR and
the excessive number of replicas required for FPR to allow 5 nodes off-line simultaneously.
PR(10) (partitioned replicas with two 10 element groups) is a good choice,
as it requires a single replica and obtains a good response time simultaneously.
Given these results, we conclude that replicas partitioned by groups are the most
advantageous alternative for NPDW if we consider both performance and flexibility
in allowing multiple nodes to be taken off-line simultaneously for maintenance
and loading reasons.
Figure 15. Response time and replicas when 5 out of 20 nodes are off-line (average
over TPC-H)
0
onl ne FPR PR( 0) PR( ) FR
nr of replicas
00:00
:
2 :0
4:
response t me (m n:sec)
N?? of Replicas
Response T me (m n:sec)
Efficient and Robust Node-Partitioned Data Warehouses 22
Copyright ?© 2007, Idea Group Inc. Copying or distributing in print or electronic forms without written permission
of Idea Group Inc. is prohibited.
Future.
Pages:
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426