Throughput versus Response Time in Cache-insensitive Scenario

Figure 9: A comparison of LRW, CSCAN, and WOW. The left panel displays achieved overall throughput versus achieved average response time. This set-up does not have much temporal locality. WOW and CSCAN have comparable performance, and both outperform LRW dramatically. Specifically, WOW increases the peak throughput over LRW by 50%. The right panel shows the target throughput corresponding to the data points in the left panel. It can be clearly seen that LRW hits an insurmountable stiff wall at a much lower throughput.

In Figure 9, we compare LRW, CSCAN, and WOW using SPC-1 Like workload directed to Full Backend on RAID-5. We use an NVS size of 4K pages each of 4KB. Hence, NVS to backing store ratio is very low, namely, 0.011%, constituting a cache-insensitive scenario.

We vary the target throughput of SPC-1 Like from 100 IOPS to 1100 IOPS. At each target throughput, we allow a settling time of 10 mins, after which we record average response time over a period of 5 minutes.

It can be clearly seen that WOW and CSCAN have virtually identical performances, and both significantly outperform LRW. In particular, it can be seen that LRW finds it impossible to support throughput beyond 515 IOPS. Demanding a target throughput higher than this point does not yield any further improvements, but rather worsens the response times dramatically. In contrast, WOW and CSCAN saturate, respectively, at 774 and 765 IOPS. In other words, WOW delivers a peak throughput that is 50% higher than LRW, and virtually identical to CSCAN.