Straggling Task Detection Improvement

[zhangpengshan]

So far, if a task/container run over threshold three times, it will be killed and fail-over make this task run again in other machine. But I found in our cluster(very busy), sometimes there are always a slow task blocking all other tasks. A good detection improvement is needed to detect such kind of task.

Not to let user set the threshold, while collecting metrics each iteration from all workers, if someone is over standard deviation too more, kill it.