Check out the new USENIX Web site.

Capacity Changes

Storage capacity can change dramatically due to workload changes or array accesses by uncontrolled hosts external to PARDA. We have already demonstrated in Section 5.3.2 that our approach is able to absorb any spare capacity that becomes available. To test the ability of the control algorithm to handle decreases in capacity, we conducted an experiment starting with the first five hosts from the previous experiment. At time $ t$ = 230 s, we introduce a sixth host that is not under PARDA control. This uncontrolled host runs a Windows Server 2003 VM issuing 16 KB random reads to a 16 GB virtual disk located on the same LUN as the others.

With $ \cal {L}$ = 30 ms and a share ratio of $ 2:2:1:1:1$ for the PARDA-managed hosts, Figure 11 plots the usual metrics over time. At $ t$ = 230 s, the uncontrolled external host starts, thereby reducing available capacity for the five controlled hosts. The results indicate that as capacity changes, the hosts under control adjust their window sizes in proportion to their shares, and observe latencies close to $ \cal {L}$ .

Figure 11: Capacity Fluctuation. Uncontrolled external host added at $ t$ = 230 s. PARDA-controlled hosts converge to new window sizes.




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

Ajay Gulati 2009-01-14