On the computational efficiency of subgradient methods: a case study with Lagrangian bounds