Towards Systematic Evaluation of Logical Reasoning Ability of Large Language Models