Throughput versus Response Time in Cache-sensitive Scenario

Figure 10: A comparison of LRW, CSCAN, and WOW. The left panel displays achieved overall throughput versus achieved average response time. This set-up has significant temporal locality since the ratio of NVS size to the size of the backend is relatively high (4.52%). WOW increases the peak throughput over LRW by 9% and over CSCAN by 129%. The right panel shows the target throughput corresponding to the data points in the left panel. It can be clearly seen that CSCAN hits an insurmountable stiff wall at a much lower throughput.

In Figure 10, we compare LRW, CSCAN, and WOW using SPC-1 Like workload directed to Partial Backend on RAID-5. We use an NVS size of 40K pages each of 4KB. Hence, NVS to backing store ratio is relatively large, namely, 4.52%, constituting a cache-sensitive scenario.

We vary the target throughput of SPC-1 Like from 300 IOPS to 3000 IOPS. At each target throughput, we allow a settling time of 10 mins, after which we record average response time over a period of 8 minutes.

It can be clearly seen that WOW dramatically outperforms CSCAN and even outperforms LRW. In particular, it can be seen that CSCAN finds it impossible to support throughput beyond 1070 IOPS. In contrast, WOW and LRW saturate, respectively, at 2453 and 2244 IOPS. In other words, WOW delivers a peak throughput that is 129% higher than CSCAN, and 9% higher than LRW.

Remark 6.1 (backwards bending)   Observe that in Figures 9 and 10 when trying to increase the target throughput beyond what the algorithms can support, the throughput actually drops due to increased lock and resource contention. This ``backwards bending'' phenomenon is well known in traffic control and congestion where excess traffic lowers throughput and increases average response time.