Can Perplexity Reflect Large Language Model's Ability in Long Text Understanding?