4. 2. Effect on neighboring Wi-Fi networks

High cumulative channel occupancy transmissions.

Po WiFi leverages the inherent fairness of the Wi-Fi Medium Access Control to ensure that it is fair to other Wi-Fi networks. As a worst-case evaluation, we consider a situation where PoWiFi always tries to achieve high cumulative channel occupancies at all times. To do this, we place our PoWiFi router in the vicinity of a neighboring Wi-Fi router–client pair operating on channel 1. We configure the Po WiFi router to transmit power packets at the highest achievable channel occupancies using our algorithm on all three nonoverlapping channels. We run iperf with UDP traffic on the neighboring router–client pair at the highest data rate and measure the achievable throughput as before. We repeat the experiments for different Wi-Fi bit rates at the neighboring Wi-Fi router–client pair. We compare three schemes: BlindUDP where our router transmits UDP packets at 1Mbps, EqualShare where we set our router to transmit the UDP packets at the same Wi-Fi bit rate as the neighboring router– client pair, and finally Po WiFi. EqualShare provides a baseline when every router in the network gets an equal share of the wireless medium.

Figure 7a shows the throughput for the three schemes, averaged across five runs. As expected, BlindUDP significantly degrades the neighboring router–client performance. Further, this deterioration is more pronounced at the higher bit rates. With Po WiFi, however, the throughput achieved at the neighboring router–client pair is higher than EqualShare. This is because Po WiFi transmits power packets at 54 Mbps; transmissions at such high rates occupy the channel for a smaller duration than, say, a neighboring router transmitting at 16Mbps. This property means that Po WiFi provides better than equal-share fairness to other Wi-Fi networks. We note that while our experiments are with 802.11g, Po WiFi’s power packets use the highest bit rate available for Wi-Fi. Thus, the above fairness property would hold true even with 802.11n/ac.

Rectifier-aware power transmissions. Next, we evaluate
the potential of our rectifier-aware technique, to signifi-
cantly reduce the average channel occupancy of the power
transmissions, while efficiently delivering power to the sen-
sors. To do this, we place our battery-free temperature sen-
each run, we run five sequential copies of iperf, 3 s apart, and
compute the achievable throughput over 500ms intervals,
with all the schemes described above.

Figure 5b plots CDFs of the measured throughput values across all the experiments. The plot shows that BlindUDP significantly degrades TCP throughput. As before, since NoQueue does not prioritize the client traffic over the power packets, it roughly halves the achievable throughput. Po WiFi sometimes achieves higher throughput than the baseline. This is due to changes in channel conditions that occur during the 3-h experiment period. The general trend however points to the conclusion that PoWiFi does not have a noticeable effect on TCP throughput at the client.

Figure 6b plots the CDFs of the channel occupancies for Po WiFi during the above experiments. The figure shows that PoWiFi has a mean cumulative occupancy of 100.9% and hence can efficiently deliver power.

Effect on PLT. We develop a test harness that uses the PhantomJS headless browser11 to download the front pages of the 10 most popular websites in the US1 100× each. We clear the cache and pause for 1 s in between page loads. The traffic is recorded with tcpdump and analyzed offline to determine PLT and channel occupancy. The router uses the default rate adaptation to modify its Wi-Fi bit rate. The experiments were performed during a busy weekday at UW CSE over a 2-h duration.

Figure 5c shows that BlindUDP significantly increases the PLT. This is expected because the 1 Mbps power traffic occupies a much larger fraction of the medium and hence increases packet delays to clients. NoQueue improves PLT over BlindUDP, with an average delay of 294ms over the baseline. PoWiFi further minimizes the delay to 101ms, averaged across websites. This residual delay is due to the computational overhead of PoWiFi from the per-packet checks performed by the kernel. This slows down all the processes in the OS and hence results in additional delays. However, increasing processing power and moving these checks to hardware can further reduce these delays. In our home deployments (Section 6), users did not perceive any noticeable effects on their web performance.

For completeness, we plot the CDFs of channel occupancies for Po WiFi in Figure 6c. The plot shows the same trend as before, with a mean cumulative occupancy of 87.6%.

0 2 4 6 8 10 12 14 16

0 10 20 30 40 50 Ach ie ved thr oug hpu t ( Mbp s )

Wi-Fi bit rate (Mbps)
EqualShare
PoWiFi
BlindUDP

0 0.2 0.4 0.6

0.8 1

0 5 10 15 20 25
Cumu
la
ti
ve
f
ra
c
ti
on
TCP throughput (Mbps)
Rectifier aware
PoWiFi
Baseline

0 0.2 0.4 0.6

0.8 1

0 5 10 15 20 25 30 35 40 45 50
C
umul
a
ti
ve
fr
a
c
tio
n
Per channel occupancy (%)
Rectifier aware
PoWiFi

0 5 10 15 20

136 TC P t hro ughpu t (M bps )

Po WiFi transmitters

(a) (b) (c) (d)

Figure 7. Effect of Po WiFi, rectifier aware and concurrent power transmissions on neighboring Wi-Fi networks. The plots show that Po WiFi power transmissions provide better than EqualShare throughput performance. Rectifier aware power transmissions further improve the throughput by reducing the per channel occupancy by a factor of 10. Additionally, increasing the number of concurrently transmitting Po WiFi devices does not degrade the performance of neighboring Wi-Fi devices. (a) Po WiFi bit-rates, (b) Rectifier aware throughput, (c) Rectifier aware occupancies, and (d) Concurrent transmissions.