Understanding the SIMD Efficiency of Graph Traversal on GPU