>
> Dual Pentium III 500Mhz, 256M RAM, 4x37.5GB IBM DeskStar disks.
> One of the disks contain one big 36GB XFS filesystem (created
> with default values of mkfs.xfs).
>
> [root@pcrd18 /root]# xfs_info /shift/pcrd18/data02
> meta-data=/shift/pcrd18/data02 isize=256 agcount=35, agsize=261630 blks
> data = bsize=4096 blocks=9157042, imaxpct=25
> = sunit=0 swidth=0 blks, unwritten=0
> naming =version 2 bsize=4096
> log =internal bsize=4096 blocks=1200
> realtime =none extsz=65536 blocks=0, rtextents=0
Just an aside here, performance will improve by using a bigger log and
some mount options.
mkfs -t xfs -f -l size=16000b /dev/xxx
would create a 16000 x 4K block log for instance. In heavy metadata update
situations, a small log tends to lead to new operations getting blocked
waiting for log space, log space being freed by flushing previously
modified metadata. Under this load you will always be log bound, but a
larger log helps some.
For mount options I would recommend
logbufs=4,logbsize=32768,kio
logbufs is the number of internal buffers used for writing transactions
into, logbsize is the size of these buffers. These two values are currently
the maximums for Linux. kio only applies for scsi at the moment, it will
reduce system overhead in getting I/O requests to the driver.
Steve
>
> > The vnode crash is being worked on, all I have managed to do so
> > far is make it harder to hit. It is basically top of the list at
> > the moment.
>
> I'll update my CVS repository soon to see the changes (I'm using
> the cvs20000829 snapshot).
>
> Peter
>
> --
> .+'''+. .+'''+. .+'''+. .+'''+. .+''
> Kelemen Péter / \ / \ Peter.Kelemen@xxxxxxx
> .+' `+...+' `+...+' `+...+' `+...+'
|