Disk to Tape Backup performance improvement

[Sundar_7]

We have a L-2000 running 11.0 connected to MSL5060L1 tape library that has 4 LTO-1 drives daisy-chained using a single onboard SCSI ( C896 Fast Wide LVD).

The intial plan was to connect the Tape libary to the A6829A Ultra160 HBA but then due to reasons the system intermittently crashed if the library is connected to the A6829A.

The backup process is to establish the BCV devices to the production std volumes, split, import and activate the VG , mount the filesystems and backup to the tape using Legato Networker.

The questions I have

1) The disk to tape backup speed is at 8-10MB/S. Is there any tweaking that we can do in the Networker that can improve the performance.

2) Has anyone faced similar problems with A6829A ? If I get A6829A working, connecting the tape libary to the A6829A will improve the performance ??

3) With Online-JFS what are the recommended mount options for the filesystems that can improve the throughput to the drive mechanism.

I would appreciate any useful inputs.

Thanks,

Sundar.

[Steven E. Protter]

I've twice backed out on posting here, because I'm not sure its useful.

We have Ultrium 230e drives hung of C896 LVD SCSI cards. We're getting nearly the rated throughput on those puppies.

There is no reason why you shouldn't be getting excellent performance off those LTO drives.

I'd start with the basics, June or September 2003 pactch set. Then I'd paruse through the LVD and SCSI patches that can be searched from the indidvidual patches clidk item on the itrc main page.

I'll assume you do have cabling and termination right, reminding that newer tape drives are self terminating. They tend to cause problems and crashes if you put a terminator on them. Could that have caused your issue on the 160 Card? I don't know. It might have not been a supported drive/card combination, only further testing will determine that.

2) No.
3)VXFS is VXFS, I don't see how its going to help.

You might want to look into kernel parameters with regards to buffers and the like.


In case there is a bottleneck, I'm enclosing some performance data collection scripts. Run these puppies while the backup is running and then look the data over for problems. Post it back for analysis in this thread.

See attachment

SEP

[Leif Halvarsson_2]

Hi,
A very common performance problem when using Ultrium drives is the disk/filesystem. Many disk/filesystems is simply not fast enough to stream a such drive particulary if the average filesize is small.

Some suggestions (I am not sure if it is supported in NetWorker but I think it is):

- Try a backup to a NULL device (/dev/null).
- Do a rawdisk backup.

Do you get better throughput with any of this tests.

[Ted Ellis_2]

have you checked with HP about your set-up? We have a SureStore HP Library here with 4 LTO-1 drives. I was specifically told not to daisy chain these drives across a single controller. They said this could overdrive the SCSI channel at the card and cause problems. I have a dedicated card for each Tape drive....

We don't use Legato, so I don't have anything specific there. I have seen discussions about adjusting the block size for the backup as being an area to test....

that said, my fastest backup tends to run 10-12 MB/s (N-Class server with Omniback) to tape.

Ted

[Michael Steele_2]

Well, here's some advice to follow for faster backups.

Never use software compression together with hardware compression. The combo can create something akin to sparse files and double your backup time.

Contiguous data in a file system backs up faster than fragmented. So think about defragging.

Because contiguous data is faster to backup first backing up to a disk and then to a tape device has the potential for being faster. This is because data being written into a new file system will be contiguous. So 'newfs' the temporary file system being used for disk backups. 'Newfs' takes a second to run and will guarantee contiguousness.

Some backup applications like Netbackup, so perhaps Legato, build a LIMIT_BANDWIDTH environment variable into their backups. Check for this possibility. May you're throttling yourself.

Raw data transfers are faster then file system transfers because they bypass the file system.

Incremental are faster to backup but slower to restore.

Full are slower to backup but faster to restore.

Finally, I've never like daisy chaining tape devices through one SCSI controller. So look into this. You'll get faster performance if you have a one to one relationship between tape drives and SCSI controller.

But if the tape library is caching data as it comes in then the daisy chaining is a non issue.

This is a good transition into buffer size. Tapes don't get written to until the buffer fills up. So if the buffer is too big then the tape drive will sit idle. So consider reducing the buffer size for faster tape device writing.

Run these basic performance commands during your backup and attach the results.

Thanks!

sar -u 5 5
sar -v 5 5
sar -d 5 5
sar -b 5 5
vmstat 5 5
swapinfo -tam

Hope this helps.