What if there are reads in the workload?

We have now seen the benefits of range writes in synthetic workloads consisting purely of writes. We now include reads in the workload, and show the diminishing benefit of range writes in read-dominated workloads. Figure 7 plots the results of our experiments.

Figure: Mixing in Reads. The graphs plot the performance of range writes to random destinations when there is some percentage of reads mixed in. We utilize track-sized ranges and write randomly to locations within a block group. The x-axis varies the percent of reads from 0% to 100%, and the y-axis plots the average time per operation (read or write). Finally, the graph on the left has 1 outstanding request to disk, whereas the graph on the right has 16.

0.9

From the figures, we observe the expected result that with an increasing percentage of reads, the benefit of range writes diminishes. However, we can also see that for many interesting points in the read/write mix, range writes could be useful. With a small number of outstanding requests to the disk and a reasonable percentage of writes, range writes decrease positioning time noticeably.