Non-Uniform Workloads

Figure 10: Non-Uniform Workloads. PARDA control with $\cal {L}$ = 40 ms. Six hosts run mixed workloads, with $\beta$ values $2:1:2:1:2:1$ .

(a) Window Size	(b) Latency (ms)	(c) Throughput (IOPS)

To test PARDA and its robustness with mixed workloads, we ran very different workload patterns at the same time from our six hosts. Table 5 presents the uncontrolled case.

Table 5: Uncontrolled access by mixed workloads from six hosts.

Host	Size	Read	Random	IOPS	Latency (ms)
1	4K	100%	100%	610	51
2	8K	50%	0%	660	48
3	8K	100%	100%	630	50
4	8K	67%	60%	670	47
5	16K	100%	100%	490	65
6	16K	75%	70%	520	60

Next, we enable PARDA with $\cal {L}$ = 40 ms, and assign shares in a $2:1:2:1:2:1$ ratio for hosts 1 through 6 respectively, plotted in Figure 10. Window sizes are differentiated between hosts with different shares. Hosts with more shares reach a window size of 32 (the upper bound, $w_{max}$ ) and remain there. Other hosts have window sizes close to 19. The average latency observed by the hosts remains close to $\cal {L}$ , as shown in Figure 10(b). The throughput observed by hosts follows roughly the same pattern as window sizes, but is not always proportional because of array scheduling and block placement issues. We saw similar adaptation in window sizes and latency when we repeated this experiment using $\cal {L}$ = 30 ms (plots omitted due to space constraints).