Max Throughput 4000 Rate 1 Hop 2 Hops 3 Hops 3500: 1890 445 21634 817 545 5.5 13 1671 1500F3 Table 2: Theoretical loss-free maximum through put over one, two, and three hops for each 802.11b 10 transmit bit-rate, with 1500-byte packets Hops Number ThroughputLatency 2000 Distance(meters 552 379 37 Avg: 2.3 Total: 33 vg:1395Avg:22 0≌ Table 3: Average TCP throughput and round-trip 法: ping latency to the 33 non-gateway nodes from each node's chosen gateway, arranged by hop-cour Even at four hops, the average throughput is com- parable to many DSL links(multi-hop TCP) Most roofnet users talk only to the Internet gateway with the best metric, and thus use routes with fewer hops than the average of the all-pairs routes. Table 3 shows the TCP 3: Link throughput versus distance for all throughput to each node from its chosen gateway, again ar nd for the links used by Srcr(bottom) nged by hop-count. The maximum hop-count is only five e most use of short high-throughput because no node is very far from the nearest gateway. The average throughput for each hop-count is typically higher hop TCP, multi-hop TCP) because the roofnet gateways happen to be more centrally located than the average roofnet node. Even at four ho throughputs of two megabits/second or more, and a few he average of 379 kbits/second is comparable to many DSL longer high-throughput links The lower graph shows just the links that Srer uses in The tables also show round-trip latencies for 84-byte ping some route. Srcr uses almost all of the links faster than packets to estimate interactive delay on a relatively idle net. two megabits/second, but largely ignores the majority of ork. Latency is affected by per-hop processing time as well the links. which are slower than that. This means that as by 802.11 retransmissions and back-offs when packets are Srer effectively favors short links of a few hundred meters, ceptable over a few hops but would be bothersome over nine goring many links that would carry packets best policy for example, four 250-meter hops that individually run at gateways, which is hardly noticeable in an interactive ses- three megabits/second yield a route with a throughput 750 kbits/ second, which is faster than most of the single 1000-meter links 3.3 Link Quality and distance A links throughput is determined by its best transmit bit- ate and the delivery probability at that bit-rate. Figure 4 While high quality 802.11 links can be constructed using hows the Cdf of delivery probabilities for the links used directional antennas, it is not clear what useful ranges and by Srer at the bit-rate chosen by SampleRate. The mediar speeds to expect with omni-directional antennas, or what delivery probability is 0.8, and nearly a quarter of the links kinds of links will be most useful to the routing protocol have loss rates of 50% or more. Section 2.5 justifies the use The upper graph in Figure 3 shows the throughput and of links and bit-rates with significant loss rates, as opposed distance of each available link. Most of the available links to favoring low-loss links. 802.11 detects the losses with it e between 500 and 1300 meters long, and can transfer ACK mechanism and re-sends the packets; this decreases about 500 kbits/ second at their best bit-rate. There are throughput but has little perceptible effect on latency, since also a small number of links a few hundred meters long with the retransmissions occur within a few millisecondsMax Throughput (kbits/sec) Rate 1 Hop 2 Hops 3 Hops 1 890 445 297 2 1634 817 545 5.5 3435 1718 1145 11 5013 2506 1671 Table 2: Theoretical loss-free maximum through￾put over one, two, and three hops for each 802.11b transmit bit-rate, with 1500-byte packets. Hops Number Throughput Latency of nodes (kbits/sec) (ms) 1 12 2752 9 2 8 940 19 3 5 552 27 4 7 379 43 5 1 89 37 Avg: 2.3 Total: 33 Avg: 1395 Avg: 22 Table 3: Average TCP throughput and round-trip ping latency to the 33 non-gateway nodes from each node’s chosen gateway, arranged by hop-count. Even at four hops, the average throughput is com￾parable to many DSL links. (multi-hop TCP) Most Roofnet users talk only to the Internet gateway with the best metric, and thus use routes with fewer hops than the average of the all-pairs routes. Table 3 shows the TCP throughput to each node from its chosen gateway, again ar￾ranged by hop-count. The maximum hop-count is only five because no node is very far from the nearest gateway. The average throughput for each hop-count is typically higher because the Roofnet gateways happen to be more centrally located than the average Roofnet node. Even at four hops, the average of 379 kbits/second is comparable to many DSL links. The tables also show round-trip latencies for 84-byte ping packets to estimate interactive delay on a relatively idle net￾work. Latency is affected by per-hop processing time as well as by 802.11 retransmissions and back-offs when packets are lost. Tables 1 and 3 suggest that interactive latency is ac￾ceptable over a few hops but would be bothersome over nine hops. Roofnet users see on average 22 ms of latency to the gateways, which is hardly noticeable in an interactive ses￾sion. 3.3 Link Quality and Distance While high quality 802.11 links can be constructed using directional antennas, it is not clear what useful ranges and speeds to expect with omni-directional antennas, or what kinds of links will be most useful to the routing protocol. The upper graph in Figure 3 shows the throughput and distance of each available link. Most of the available links are between 500 and 1300 meters long, and can transfer about 500 kbits/second at their best bit-rate. There are also a small number of links a few hundred meters long with 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 Throughput (kbits/sec) Distance (meters) 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 Throughput (kbits/sec) Distance (meters) Figure 3: Link throughput versus distance for all links (top) and for the links used by Srcr (bottom). Srcr makes the most use of short high-throughput links. (single-hop TCP, multi-hop TCP) throughputs of two megabits/second or more, and a few longer high-throughput links. The lower graph shows just the links that Srcr uses in some route. Srcr uses almost all of the links faster than two megabits/second, but largely ignores the majority of the links, which are slower than that. This means that Srcr effectively favors short links of a few hundred meters, ignoring many links that would carry packets a kilometer or more in one hop. Fast short hops are the best policy: for example, four 250-meter hops that individually run at three megabits/second yield a route with a throughput of 750 kbits/second, which is faster than most of the single 1000-meter links. A link’s throughput is determined by its best transmit bit￾rate and the delivery probability at that bit-rate. Figure 4 shows the CDF of delivery probabilities for the links used by Srcr at the bit-rate chosen by SampleRate. The median delivery probability is 0.8, and nearly a quarter of the links have loss rates of 50% or more. Section 2.5 justifies the use of links and bit-rates with significant loss rates, as opposed to favoring low-loss links. 802.11 detects the losses with its ACK mechanism and re-sends the packets; this decreases throughput but has little perceptible effect on latency, since the retransmissions occur within a few milliseconds