Reliable and Fault-Tolerant Storage Systems - Scalable Continuous Media Streaming Systems - page 102

Figure 5.8 A snapshot of track retrievals at time t = l i with Active Disk Synchronization

Theoretically, with ADS the deviation-bound D asyn will become zero. In practice, small
deviations might still exist because the server is likely to send disk commands serially to each
of the surviving disks. Assuming this deviation is small compared to

, then

D asyn

= 1 for D asyn ≤ (5.44)

and the buffer requirement is reduced to

B r = (2 N D − 2) SY max (5.45)

5.7 Performance Evaluation
Using the performance models derived in the previous sections, we present in this section
numerical results computed for five disk drive models to quantitatively compare the studied
algorithms. The disks' parameters are extracted from the disk specifications inGanger et al . [21]
and summarized in Table 5.2. Unless stated otherwise, the results are computed using a disk
array configuration of four data disks, one parity disk, and one spare disk.

5.7.1 Comparison of Rebuild Time
Figures 5.9 and 5.10 show the rebuild time versus server utilization for block-based rebuild
and track-based rebuild respectively. We observe that the rebuild time increases modestly
until around a utilization of 0.8, after which it increases rapidly due to the limited capacity
available for rebuild. For example, rebuild time for the QuantumAtlas-10K disk increases from
44.4 minutes at 9 for block-based rebuild. Comparing

Figure 5.9 with Figure 5.10, it is clear that track-based rebuild significantly outperforms block-
based rebuild. With the same disk model, the rebuild time for track-based rebuild is only
12.8 minutes at

ρ = 0 . 5 to 221.9 minutes at ρ = 0 .

9.
This result is encouraging, as rebuilding a failed disk requires less than 1.5 hours even at a
server utilization of 0.9. Given that a service provider is likely to dimension a system to operate

ρ = 0 . 5 and 87.7 minutes at ρ = 0 .