Algorithms and hardness results for parallel large margin learning