Flat Delivery

  In the flat delivery model, all sources transmit their media over a single network channel that has a uniform aggregate rate. We assume either that the multicast session has a fixed bandwidth allocation or that a sender-based rate-adaptation algorithm computes a rate limit for a given source. In either case, we call this overall bandwidth limit the session bandwidth.

The flat delivery adjustment decision maps the transmission weight directly to the source transmission bandwidth, i.e., if is the transmission rate for source R₀...R_M-1, then we set

B_k = w_kB

where B is the session bandwidth, and w_k is the average source weight for S_k.

More generally, the mechanism that maps the receiver interest vector into a source-rate control decision is arbitrary and can be customized for the application or environment at hand. Having recognized this, we adopt the following rule (assumed for the remainder of the paper): 95% of the session bandwidth is allocated across sources that each have at least one interested receiver (i.e., those whose average weight is non-zero) while the remaining 5% is equally apportioned over the remaining sources (i.e., those whose average weight is zero). As a result, each active source multicasts its signal continuously, though perhaps at a very low rate. Each receiver can then build a thumbnail display of all the active sources, thereby enhancing the ``sense of presence'' of a multimedia collaboration. The following pseudo-code segment implements this policy:

if (avg_weight > 0) {
	share = 0.95 * avg_weight
} else {
	share = 0.05 / zero_weight_src_num
}
tx_rate = session_bandwidth * share

To further illustrate the flat delivery variant of SCUBA, we present the following example. Let S₀ and S₁ be senders and let R₀, R₁ and R₂ be receivers. Assume that at a certain time, the source weights of {S₀,S₁} advertised by each receiver are:

• R₀: {0.8, 0.2}
• R₁: {0.8, 0.2}
• R₂: {0.5, 0.5}

Upon receipt of these reports (for simplicity we assume the sample set includes all three receivers), the average source weights computed by S₀ and S₁ become 0.7 and 0.3, respectively. Assuming the session bandwidth, B, is 128 kb/s, proportional bandwidth allocation results in S₀ transmitting at 128 * 0.95 * 0.7 = 85.12 kb/s and S₁ at 36.48 kb/s.

To better illustrate the dynamics of this distributed algorithm, we implemented it in the ns network simulator [15] and ran a simple experiment consisting of four sources and three receivers across a single bottleneck link. All sources were placed on one side of the link while the receivers were placed on the other. The session bandwidth was set to 400 kb/s.

  

Figure 3: Simulation of the flat-delivery variant of SCUBA

Figure 3 shows the resulting sample paths. The sources join the session at times 1.4, 1.7, 2.0 and 2.3, respectively. Before time 3, no receivers report interest, and thus the sources equally share the 5% reserve. Then, at times 3.0, 3.2, and 3.3, receivers R₀ R₂ and R₃ each respectively announce interest in source S₁. In response, S₁. transmits at the full 95% allocation (i.e., 380 kb/s). Finally, at time 3.4, R₂ shifts its interest from S₁ to S₃ and from that point on S₁ and S₃ share the 95% allocation in proper proportion.