Information Technology Reference
In-Depth Information
100
80
60
40
20
0
0
0.2
0.4
0.6
0.8
1
Server Utilization
Quantum Atlas-III
Quantum Atlas-10K
Seagate Barracuda
Seagate Cheetah
IBM 9ES
Figure 5.11 Comparison of rebuild time reduction by track-based rebuild ( Q = 64KB, N D = 5)
rebuild improves. However, even at a very large block size of 640KB, track-based rebuild still
outperforms block-based rebuild by about 30%.
5.7.4 Buffer Requirement
We plot the buffer requirement for the studied algorithms versus number of disks in Figure 5.14
for the Quantum Atlas-10K disk model. We observe that track-based rebuild without buffer
sharing has the largest buffer requirement as expected. However, even with buffer sharing
track-based rebuild still requires more buffers than block-based rebuild. This is due to the fact
that the block size (64KB) used is smaller than the track size (varies from 114.5KB to 167KB)
and hence the rebuild buffers dominate the buffer requirement.
By contrast, the pipelined rebuild algorithmhas only a slightly larger buffer requirement than
the best scheme - block-based rebuild with buffer sharing. For a five-disk server, pipelined
rebuild requires only 0.7MB to 1.5MB more buffers than block-based rebuild with buffer
sharing (see Table 5.3). Note that block-based rebuild with buffer sharing is already optimal
because the same server will require just as much buffer without the rebuild option. Therefore,
with pipelined rebuild, we can achieve a significant gain in rebuild performance through track
rebuild and at the same time avoid the large buffer requirement.
Search WWH ::




Custom Search