SEO

Hi there,

Most of my systems use Oracle 9, AIX 5.2 & external (EMC) subsystems.

Up to now, I have been creating JFS2 inline-logged filesystems with CIO.
I have however found a document (Tuning AIX commands for JFS2 by Jaqui
Lynch) wherein it is stated: "You’ll have performance
problems with a high write workload if you use inline logs."

Considering the fact that I cannot practically predict the placement of
data on disks / nr. of spindles (since the logical hdisks I am presented
do not necessarily equate to physical disks), what is the correct policy
- inline JFS2 or not?

Can anyone here comment on what best practice should be?

0xbeef

The best thing to do might to test both in your envionment.
Perhaps you could install Bonnie and test both types of file systems.

I would reccomend that every filesystem get its own log - INLINE or
external.
For filesystems that you think will get a lot of IO, use an external
log.

I have been using dedicated JFS logs for every major filesystem that I
think will have a lot of IO.

If you're concerned, then just use a dedicate JFS log. It only takes
one PP.

FYI:

root@unxp1:/home>mount
node mounted mounted over vfs date
options
-------- --------------- --------------- ------ ------------
---------------
/dev/hd4 / jfs2 Mar 28 11:06
rw,log=/dev/hd8
/dev/hd2 /usr jfs2 Mar 28 11:06
rw,log=/dev/hd8
/dev/hd9var /var jfs2 Mar 28 11:06
rw,log=/dev/hd8
/dev/hd3 /tmp jfs2 Mar 28 11:06
rw,log=/dev/hd8
/dev/hd1 /home jfs2 Mar 28 11:06
rw,log=/dev/hd8
/proc /proc procfs Mar 28 11:06 rw
/dev/hd10opt /opt jfs2 Mar 28 11:06
rw,log=/dev/hd8
/dev/sybase_lv /home/sybase jfs2 Mar 28 11:06
rw,log=INLINE
/dev/Ptempdb_dat_lv /home/Ptempdb_DAT jfs2 Mar 28 11:06
rw,log=/dev/Ptempdb_dat_log
/dev/Ptempdb_log_lv /home/Ptempdb_LOG jfs2 Mar 28 11:06
rw,log=/dev/Ptempdb_log_log
/dev/Pworkdb_dat_lv /home/Pworkdb_DAT jfs2 Mar 28 11:06
rw,log=/dev/Pworkdb_dat_log
/dev/Pworkdb_log_lv /home/Pworkdb_LOG jfs2 Mar 28 11:06
rw,log=/dev/Pworkdb_log_log
/dev/Dtempdb_lv /home/Dtempdb jfs2 Mar 28 11:06
rw,log=/dev/Dtempdb_log
/dev/Dworkdb_lv /home/Dworkdb jfs2 Mar 28 11:06
rw,log=/dev/Dworkdb_log


Miles

0xbeef wrote:
> Can anyone here comment on what best practice should be?

Performance Problems with the JFS log tend to come from filesystems
where a lot of filesystem metadata changes happen. Metadata changes come
from creating/deleting/renaming files and form changing the file size.
If you have the log inlined then the disk head must travels from the
data area of the disk to the log area. Disk head movements are slow !
If you have several disks with filesystems sharing a single log you have
a good chance that the log and the filesystem are on different disks
and disk heads move not that much. We've dedicated a disk for the JFS
log for some specific workloads.
However, you say the your data is an oracle database. The database files
change (from the filesystem point of view) quite rarely. The most active
files are the logs and I doubt you'll have a log switch a couple of
times a second. The other thing is that you are writing to a SAN disk
controller. These controllers do mitigate the problem a log as they can
write to (protected) cache and don't need to write to disk for every
write request.
I don't think you'll see any difference in performance between inlined
or not inlined for your application.

Markus

Miles wrote:
> If you're concerned, then just use a dedicate JFS log. It only takes one PP.

Correct, but if your PP size is 128MB or higher, then with only eight
JFS logs, you have used up 1GB of space (2GB if the jfslog is
mirrored).

We use one jfslog area for several file systems so as not to end up
with this scenario. Our guideline is similar to Markus's...if it will
be a busy jfs (for I/O) or if the jfs size exceeds 512 PPs, we dedicate
a jfslog to that area.

Steve