We now turn to more complex traffic patterns as opposed to the throughput of 1-minute TCP flows. Corporate network traffic can be modeled as a series of short flows arriving at various times [12]. The metric of interest for such traffic patterns is the flow completion time.
To compare the performance of DenseAP system with corporate WLAN with such on/off traffic, we carry out the following experiment. We use 12 clients, all of which are active simultaneously. Each client downloads 2000 files from a central server on the corporate network. The sizes of files are chosen from a Pareto distribution with mean of 125KB and shape parameter of 1.5. The time between start of successive downloads is chosen from an exponential distribution with a given mean. By changing the mean time between successive requests, we can control the amount of offered load generated by each client.
We consider three scenarios. In the first scenario, the mean interarrival time between successive downloads is 0.5 seconds. This corresponds to a mean offered load of about 2Mbps per client. In Figure 8, we see that when 12 clients are simultaneously active on the corporate WLAN, the median per-client throughput is 750Kbps. Thus, the 2Mbps offered load represents a heavy overload of the corporate network. We similarly construct a fully loaded, and highly loaded scenario using mean interarrival times of 1 and 2 seconds, respectively. The details are shown in Table 1.
We repeat the experiment on corporate WLAN, as well as the DenseAP system using 24 APs and 8 channels. The median flow completion times under corporate WLAN, and DenseAP system are shown in Table 1. We see that the median flow completion time for corporate WLAN is very high under the overload and full load scenarios. These high flow completion times are detrimental to user experience. On the other hand, in DenseAP the median flow completion time is essentially equal in all three cases, since the load on the system is substantially lower than its capacity.
NSDI-2008
