And the most important, i found that with the first method, our cluster nodes drift more than 15 seconds frequently. Thanks to
Linux ntpd and ntpdate service
now i know the reason, maybe for some reason the timedrift is bigger than 2 seconds and ntpd refused to do anything then.
For what David Schwartz said, i found this useful:
As mentioned elsewhere, NTP provides a smooth time correction. If the applications on your server don't mind have whole seconds go missing, or doing the same seconds over again, then ntpd doesn't gain you much over ntpdate.
If on the other hand you do have time-sensitive applications that are sensitive to seconds, or even worse are sensitive to partial seconds, then ntpd is by far the better choice. Novell eDirectory timestamps updates for update-collision handling, which becomes critical if the updates come very fast (such as during the morning login rush). A syslog server needs to have time accurate to at least the half second in order to keep sane logs.