The results show
Disk Processing N?? Instrs. Network
seek time 10 ms read page 3,000
connection
speed
(default)
100 MB/s
settle time + ctrller
delay
per access 3 ms +
1 ms per page
process
bitmap
page
1,500 send
message 1,000 + #B instructions
Seq. transfer rate
/ node 100MB/s
extract &
hash/probe
table row
250 receive
message 1,000 + #B instructions
CPU.speed 50 MIPS message size
(small) 128 B
Memory.Buffer 500MB/node message size
(large) 1 page (4 KB)
Figure 8. Basic parameters for simulation
Figure 9. Response time and speedup VS N?? of nodes (100 MB/s) (log-plots): (a)
RT vs N?? of nodes and (b) speedup vs n?? of nodes
0
00
000
0 2 0 00 2 0
n?? of nodes
response t me (secs)
0
00
000
0 2 0 00 2 0
n?? of nodes
Speedup
PRS
PFRD-H
WBP
WBP+JB
(a) (b)
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.
that the simulated response time was reasonably accurate. Although its prediction
for PRS was slightly higher than the actual response time for a number of nodes
above 100, the advantage of WBP is still very evident from these results.
Finally, we also ran the whole TPC-H query set against WBP and PRS to compare
replication vs. partitioning on the following system: 25 nodes; TPCH 50 GB; each
node with Pentium III 866 MHz CPU; 80 GB IDE hard disks; 512 MB of RAM; 100
MB/s switched network).
Pages:
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421