Handling Bursts

Earlier we showed that PARDA maintains high utilization of the array even when some hosts idle, by allowing other hosts to increase their window sizes. However, if one or more VMs become idle, the overall $\beta$ of the host must be adjusted, so that backlogged VMs on the same host don't obtain an unfair share of the current capacity. Our implementation employs the technique described in Section 3.4.

We experimented with dynamically idling one of the OLTP VM workloads running on host 1 from the previous experiment presented in Figure 12. The VM workload is stopped at $t$ = 140 s and resumed at $t$ = 310 s. Figure 13 shows that the $\beta$ value for host 1 adapts quickly to the change in the VM workload. Figure 12(a) shows that the window size begins to decrease according to the modified lower value of $\beta=4$ starting from $t$ = 140 s. By $t$ = 300 s, window sizes have converged to a $1:2$ ratio, in line with aggregate host shares. As the OLTP workload becomes active again, the dynamic increase in the $\beta$ of host 1 causes its window size to grow. This demonstrates that PARDA ensures fairness even in the presence of non-backlogged workloads, a highly-desirable property for shared storage access.

Figure 13: Handling Bursts. One OLTP workload on host 1 stops at $t$ = 140 s and restarts at $t$ = 310 s. The $\beta$ of host 1 is adjusted and window sizes are recomputed using the new $\beta$ value.