OnlineJFS performance question...

[Eric Ladner]

We are using OnlineJFS to obtain greater performance from the file system.

My question is this?

Given a disk device (/dev/dsk/c4t0d0, for example), would using LVM to configure the disk incur a performance penalty over just using the device directly?

I.e., which of the two configuration opgions below delivers higher performance? Option A or B? (for comparison purposes, both options are using mincache=direct,convosync=direct to achieve "Direct I/O" per the OnlineJFS docs)

Option A:
pvcreate
vgcreate vg01
lvcreate -l xxx vg01
newfs -F vxfs /dev/vg01/rlvol1
mount /dev/vg01/lvol1 /drive

Option B:
newfs -F vxfs /dev/dsk/c4t0d0
mount /dev/vg01/lvol1 /drive

My initial thoughts are that Option B is faster because accesses to the disk don't have to go through the LVM subsystem, but go directly to the device.

Thoughts?

[A. Clay Stephenson · Accepted solution]

Hi Eric:

Your analysis is correct but just barely so and only for single disks. The actual overhead associated with LVM raw lvols is very small. All that happens is that a request to character to major device 64 minor device 0x020001 block 3457 for example simply does a table lookup to say that that is really character major device 188 minor 0x032000 block 239 (for example). This is a very fast lookup and happens many times faster than the actual disk i/o.

If you then compare LVM raw/io where the lvol is stripped across multiple disks/controllers then LVM wins hands down. Given the ease of extending LVM lvols, stripping, and mirroring -this is really a no brainer - LVM.

The analysis is identical with block i/o devices but since you mentioned the =direct vxfs options, I thought I would answer it as raw since that is what you are emulating.

Also, you probably want to add nodatainlog,delaylog to your OnlineJFS mount options as well.


Regards, Clay

[Tim D Fulford]

If you option A with multiple disks and striping (say extent) then option A could be made faster.

e.g.
pvcreate -f /dev/rdsk/
pvcreate -f /dev/rdsk/
:
mkdir /dev/vg01
mknod /dev/vg01/group c 64 0x010000
vgcreate vg01 /dev/dsk/ /dev/dsk/
lvcreate -l xxx -D y vg01
newfs -F vxfs /dev/vg01/rlvol1
mount /dev/vg01/lvol1 /mount

You copuld get furthur performance increases if you put the disks on different SCSI controllers

Tim