Split federated learning divides a neural network between resource-constrained clients and a server. This reduces client-side computation, but every minibatch can require an exchange of intermediate features, called smashed data, and cut-layer gradients. Dynamic Split Federated Learning with a Genetic Algorithm (DSFL-GA) adapts client participation and model cuts to available resources. When every round must select exactly K clients and every selectable action carries the same payload, the communication penalty is constant and cannot influence a binary GA. Our Communication-Aware DSFL-GA (CA-DSFL-GA) extends the DSFL-GA planner with one categorical action per client: exclusion or one feasible cut. The fitness combines the reconstructed base scheduling utility with normalized planned application payload, and a fixed-quota adaptation of randomized Top-k represents the boundary tensors. We compare seven methods for 1,000 rounds using 100 logical clients, exactly 80 selections per round, and three paired seeds. Relative to the otherwise identical matched control, CA-DSFL-GA-0 (λ_C = 0), CA-DSFL-GA uses 55.84% less serialized application payload, with a 0.26-percentage-point lower mean terminal accuracy on the custom combined-MNIST held-out set. Relative to our reconstructed dense DSFL-GA baseline, it uses 74.91% less payload, with a 0.90-percentage-point lower mean terminal accuracy. These results concern logical-client emulation and application payload; they do not measure physical network traffic or device-level performance.