Effects of the Number of Segment Buffers with Real-world Traces

Figure 5: Overall write cost versus Disk Utilization.

Figure 6: Overall write cost versus Segment Sizes. Disk utilization is 95%.

Figure 5 shows the results of the overall write cost versus disk utilization for the four real-world traces. We varied the number of segment buffers of WOLF from 2 to 4. We also varied the segment buffer size of the LFS from 256 KB to 1024 KB.

Increasing the number of segment buffers in WOLF would slightly reduce the overall write cost but does not have a significant impact on the overall performance.

The reason we studied LFS with different segment buffer sizes, is to show that the performance gain of WOLF is not due to the increased buffer numbers (hence the increased total buffer size). The separated active/inactive data layout on disk segments contributes to the performance improvement. In fact, for LFS, increasing the segment buffer sizes may actually increase the overall write cost. This observation is consistent with previous studies [9,15]

Note that because WOLF uses more segment buffers than the LFS does, data may stay in RAM longer. However, this does not poses a reliability problem. As discussed before, in WOLF, if the segment buffers contain data older than 30 seconds, they will be flushed to the disk, just as LFS.