A reader may wish to compare the above proposals and ask: Is one of them truly superior? We note that encouraging results have been reported for all three protocols:
• CoopMAC: In [122], it was reported that CoopMAC provides network capacity gains ranging from 40% to 60% relative to 802.11g, with the higher figure requiring receiver combining. Improvements in channel access delay and user energy efficiency on the order of 20%-40% were also observed.
• HARBINGER: In [199], HARBINGER was found to provide a power savings ranging from 7 to 20 dB over direct transmission and from 1 to 3 dB over multihop routes. The savings increase as the number of codeword blocks increases, as one would expect [25].
• VMISO: In static networks of 200 nodes and CBR traffic, it was observed in [85] that VMISO increases throughput from 30% to 100%. VMISO also decreases average delay by roughly 50%. The largest throughput improvements occur at light loads with few sessions. Similar improvements are reported under scenarios with mobility. The authors argue that the diversity afforded by a VMISO link improves the link lifetime. The resulting savings in the overhead of route repair and discovery more than offset the overhead in setting up the VMISO links.
• COPE: Experimentation with a 20 node testbed showed that COPE performance depends on the link topology and traffic flow characteristics [91]. For example, in a congested wireless network serving UDP flows, COPE provides a throughput increase of a factor 3 to 4. In other circumstances, the improvements are less dramatic. A mesh network connected to an Internet gateway yields throughput increases of 5%-70%, with the gains increasing with the fraction of down link traffic. In general, two kinds of gains are noted: coding gains in which COPE delivers packets with fewer transmissions and MAC gains in which COPE compensates for packet dropping that would otherwise be induced by the fairness properties of the 802.11 MAC.
Head-to-head comparisons are difficult because each protocol requires evaluation in a custom environment with a mix of analytic modeling, simulation, and testbed evaluation.
Although CoopMAC could be implemented in an ad hoc network, published results have employed MAC channels with a group of stations transmitting to an access point. A key barrier is that a cooperative diversity receiver cannot be implemented without low-level access to the wireless interface. Thus, the evaluation of CoopMAC employs a custom simulator "to faithfully model all the critical MAC and PHY layer features of IEEE 802.11" [122].
Since HARBINGER makes no particular assumption on the routing algorithm, none was tested. Instead, instructive predetermined configurations of nodes and routes were used, with an emphasis on nodes or clusters of nodes along a line. The MAC protocol was not explicitly modeled and interference from other network links was approximated by noise. Finally, signaling overhead that would have been incurred by a practical routing protocol was neglected.
Using an OPNET simulation, VMISO received an extensive performance evaluation. However, symbol-level effects were modeled coarsely: for each cooperative link, each transmitter u at distance du from the receiver was characterized by distance-dependent attenuation and a Rayleigh fading variable au. The received SNR wa^u au/dU, where the sum is over the transmitters participating in the VMISO link. A packet is decoded correctly if the received SNR is above a threshold (SNRTH — D) dB where SNRTH is a SISO threshold and D is a diversity gain. Interference was neglected under the assumption that a node can use RTS/CTS to gain access to the channel while silencing potential interferers. This abstraction allowed experimentation in a network with 200 nodes and various mobility models.
COPE does not employ PHY layer cooperative signaling, a simplification that enabled a testbed implementation. Although limited to a 20 node network, the COPE testbed evaluation was perhaps the most complete. The experimental results highlighted how system performance depends on the network configuration and on the traffic induced by higher layer protocols.
Unfortunately, a unified framework for comparing cooperative protocols is unlikely to emerge. For protocols that implement PHY layer cooperation, symbol-level simulations are computationally too tedious to permit generating sufficient numbers of packets for even a single network session, much less dozens or hundreds of sessions. On the other hand, it is non-trivial to accurately abstract PHY layer effects. The modeling task becomes especially difficult if interference from simultaneous transmissions cannot be neglected.
