Thanks - will comply and send out the info.
Also we just built a profiling kernel (thanks to SGI again for
the kernprof tool). Need to run this first to see what's going on.
----- Original Message -----
From: "Nathan Scott" <nathans@xxxxxxx>
To: "Wendy Cheng" <s_wendy_cheng@xxxxxxxxxxx>
Cc: "xfs mailing list" <linux-xfs@xxxxxxxxxxx>
Sent: Tuesday, June 03, 2003 1:19 AM
Subject: Re: xfs performance tuning
| On Mon, Jun 02, 2003 at 09:29:56PM -0400, Wendy Cheng wrote:
| > Great thanks ! - but these are kind of obivous.
| Why not post the XFS and ext2 numbers you are seeing currently
| with your current filesystem geometry (esp. xfs_info output),
| mount options, and kernel build options, and then any benchmark
| variations that that the suggested options make, then it might
| be clearer where to look.
| There are plenty more knobs to tweak, yes. You should also
| check that you are not doing any extra checks (e.g. do you have
| debug or ACLs enabled?) that are not being done in ext2.
| > Are there more ? :). Could I turn off / delay
| > kupdated and assume that when incore log buffers
| > are full, it would be automatically flushed into
| > disk without kupdated's help ? We have lots
| > of memory to spare and have turned on highmem IO.
| > Wendy
| > ---------
| > ----- Original Message -----
| > From: "Nathan Scott" <nathans@xxxxxxx>
| > To: "Wendy Cheng" <s_wendy_cheng@xxxxxxxxxxx>
| > Cc: "xfs mailing list" <linux-xfs@xxxxxxxxxxx>
| > Sent: Monday, June 02, 2003 7:07 PM
| > Subject: Re: xfs performance tuning
| > | On Mon, Jun 02, 2003 at 06:15:11PM -0400, Wendy Cheng wrote:
| > | > We're trying out XFS using SPECsfs benchmark and
| > | > could only get ~ half of the IOPS when compared
| > | > with EXT2. It might not be a fair comparison since EXT2
| > | > is not a journaling file system but would love to know
| > | > any tuning parameter to boost the number. Could someone
| > | > help out ? Thanks...
| > |
| > | I would suggest starting with the mkfs options "-l size=XXX" and
| > | "-i size=XXX" (log size and inode size). Increasing the number
| > | of incore log buffers you're using ("-o logbufs=XXX" mount option)
| > | might help you too.
| > |
| > | cheers.
| > |
| > | --
| > | Nathan
| > |