On Mon, Mar 26, 2012 at 2:58 AM, Richard Elling <richard.ell...@richardelling.com> wrote: > On Mar 25, 2012, at 10:25 AM, Aubrey Li wrote: > > On Mon, Mar 26, 2012 at 12:48 AM, Richard Elling > <richard.ell...@richardelling.com> wrote: > > This is the wrong forum for general purpose performance tuning. So I won't > > continue this much farther. Notice the huge number of icsw, that is a > > bigger > > symptom than locks. > > -- richard > > > thanks anyway, lock must be a problem. the scenario here is, apache > causes a bunch of stat() > syscall, which leads to a bunch of zfs vnode access, byond the normal > read/write operation. > > > That does not explain the high icsw. In a properly sized system icsw will be > on the order of 0 to 100, closer to 0. > > If there are millions of files, then you should check the DLNC hit rate.
Involuntary context switches might be another story I need to take care of. But I'm not aware it's controllable, ? > > > The problem is, every zfs vnode access need the **same zfs root** > lock. When the number of > httpd processes and the corresponding kernel threads becomes large, > this root lock contention > becomes horrible. This situation does not occurs on linux. > > > I disagree with your conclusion and I've seen ZFS systems do millions of > stats() > per second without issue. What does prstat -Lm show? > -- richard > I have ever not seen any issues until I did a comparison with Linux. Thanks, -Aubrey _______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss