In Figure 6, we compare LRW, CSCAN, and WOW using random write workload (Section V-C) directed to Full Backend on RAID-5 and RAID-10. Since the workload has almost no temporal locality, the throughput of LRW remains constant as the NVS size increases. In contrast, WOW and CSCAN exhibit logarithmic gain in throughput as a function of the size of NVS by exploiting spatial locality (also see the related discussion in Section II-B). For RAID-10, at the lowest NVS size of 32 pages, WOW and CSCAN outperform LRW by 16%, while, quite dramatically, at the NVS size of 65,536 pages, WOW and CSCAN outperform LRW by 200%. Similarly, for RAID-5, at the lowest NVS size of 64 pages, WOW and CSCAN outperform LRW by 38%, while, quite dramatically, at the NVS size of 16,384 pages, WOW and CSCAN outperform LRW by 147%.
While, for brevity, we have shown results for a queue depth of 20. When we used a larger queue depth, performance of all three algorithms increased uniformly, producing virtually identical curves. Increasing queue depth beyond 128 in either RAID-10 or RAID-5 does not seem to help throughput significantly.