Effect of the Delayed Write Buffer Size

Figure 13: Effect on TPC-C throughput as we^M increase the size of the delayed write buffer to 100,000 blocks.^M The total number of disks is a constant (36).

We have seen that replica propagation imposes a significant cost on update-in-place-based disk arrays such as RAID-10 and SR-Array. One possible way of alleviating this burden to make these alternatives more attractive is to use a larger delayed write buffer. A larger delayed write buffer is useful in two ways. One is that it may allow larger batches of replica propagations to be scheduled and these larger batches can utilize the disk bandwidth more efficiently. The second source of efficiency is that a larger buffer can potentially more effectively smooth the burstiness so that replica propagation does not have to occur in the foreground due to lack of buffer space. Figure 13 shows the results of repeating the throughput experiments shown in Figure 11 after we have increased the delayed write buffer from 10,000 blocks to 100,000 blocks. As expected, the curves representing array configurations that require data replication have all shifted to the right, signaling higher maximum sustainable throughput rates. However, even with this aggressive delayed write buffering, the high cost of update-in-place is still apparent and the advantage of eager-writing is still significant.