Managing intermittent generation in electric power systems with high penetration of renewable sources of energy presents major operational challenges. Faster, more efficient optimization techniques are essential to mitigate this intermittency and ensure grid reliability. Convex relaxations of optimal power flow (OPF) problem offer tractable mean of solving the non-linear, non-convex OPF problem. Specifically, the semidefinite relaxation yields the tightest lower bounds for the OPF but require careful exploitation of sparsity to remain computationally viable when scaling to large problem instances. This exploitation can be achieved through clique decomposition of the semidefinite constraint. In this paper, we experiment with various clique decomposition algorithms and demonstrate that the resulting OPF solve time is highly sensitive to the choice of decomposition. Our main contribution is showing that the optimal decomposition depends on both the network topology and the demand profile. We categorize networks into two types: those with a preferred decomposition that performs well regardless of demand, and those where demand significantly impacts the optimal decomposition choice. This insight opens the possibility of using a learning-based approach to predict the best decomposition for minimizing OPF solve time, tailored to the network demand.