Our SA turned to the /var/log/message at the very beginning when sth. went wrong, but no luck.
Everything about the hardware is good.
My mistake, the profile was opened during the stress test. 
I should do the test again.
However, the slow node has turned to normal since yesterday after we isolated it for two days.
During the isolation, we did nothing special except my benchmark test. After that, we left it alone for two days and then put it into production again. Now it's completely fine, like nothing had happened.
We can't reproduce this problem any more..