On 6/28/05 6:47 PM, "Dave Kleikamp" <[EMAIL PROTECTED]> wrote:
> On Tue, 2005-06-28 at 15:42 -0700, Jeff Block wrote: >> I have another question along the lines of external journal. We were under >> the impression that there is a performance gain when using an external >> journal as opposed to writing the journal directly to a RAID 5. In some of >> our testing, we did in fact see a performance gain. > > There can be a performance gain if the journal is on it's own physical > disk. Since writes to the journal are sequential, having the journal on > it's own disk eliminates latency due to the write head moving. You can > actually have both file systems share the same journal, so that > journaling for both file systems is done sequentially. If you want to > attempt this, make sure you're using a recent version of jfsutils. I'm using jfsutils version 1.1.7, 22-Jul-2004 and the man page says that you each filesystem needs to have it's own external journal device. > > On disks where the writes are cached so that they appear instantaneous > to the writer, the performance gain may not be significant. > >> But, does this make sense? Are we trading stability for a measly >> performance gain? > > You shouldn't be giving up stability, but I need to figure out what's > happening so I can fix it. I plan to play around with md devices > tomorrow, to see if I can recreate, then solve the problem. > >> Also, is it possible to move the journal back to the local device (without a >> reformat)? > > Not unless you can extend the size of the volume. When you format the > volume with an external journal, no space in the volume is reserved for > the journal, and we don't have the tools to shrink the volume. > > In case the problem is somehow related to the md devices, you can try > putting the journal(s) on a different non-md partition. This may get > you going until I can figure out exactly what's going on. I've tried this and it has failed as well. I've removed one of the partitions from the md mirror and run mkfs on it. The mkfs doesn't finish and gives the same error as before. Again, this problem seems to only occur when making the journal device with mkfs (I can run a normal 'mkfs.jfs /dev/sd...' just fine). I've attached the strace output to 'mkfs -f -J journal_dev /dev/sdc7'. /dev/sdc7 was removed from the mirror. It doesn't show up in /proc/mounts, /proc/mdstat, and it is in /proc/partitions. Thanks. Jeff
strace_out.rtf
Description: Binary data
