Tests

We tested the rs-link software indoors, on a link through several walls, between two laptops with Prism 2 cards. The link distance was picked to be in the ``marginal zone'' where each station could receive packets from the other, but with errors.¹⁴

Unlike the measurements in Figure 4, this was a two-way test over our error-corrected link. One laptop sent 105 ICMP ping queries to the Ethernet broadcast address and tallied the responses.

The measurements were as follows:

1. 105 pings were transmitted, each with a 1,146 known contents.

2. 31 pings, or 30 percent of those transmitted, did not receive a response.

3. 8 pings, or 8 percent of those transmitted, received a perfect response.

4. 10 pings, or 10 percent of those sent, received a response that the rs-link decoder could not salvage (i.e., more than 192 octets incorrect).

5. The remaining 66 pings, or 63 percent of those sent, received a an errored response that the rs-link decoder was able to salvage.

These data are presented graphically in Figure 6. Our two-way loss rate is at the right-hand side of the figure -- 192 errored octets permitted. The two-way loss rate for normal 802.11 (without link-layer retransmissions) is at the left-hand side of the figure - no errored octets permitted.

Figure 6: Fraction of two-way pings lost as a function of octet error tolerance, in a 1,530-byte ping response. Note log-log scale.

The CPU load is tolerable on modern PCs. A 200 MHz Pentium Pro is just able to saturate a 1 Mbit/sec link using all its processing power. A 2 GHz Pentium 4 has no trouble.