Slow Backup with Ultrium 460

[Ronald W. Satz]

My leasing company replaced my previous Ultrium 460 (which was SCSI-challenged). The new one is recognized quickly in the SCSI POST. It also performs well on the HPTapePerformance program, transferring 19.61 MB/sec (which is between the min-max specs of 10 and 30 MB/sec), or 1176.6 MB/min. In SCO UNIX the backup speed is 553.5 MB/min, which is acceptable, if not stellar. (Where is the 2 GB/min?) However in Windows XP, using the native applet NTBackup, the speed is a mere 99.59 MB/min. The SCSI host adapter is the Adaptec 39320. (Both the UNIX and the Windows XP hard drives are LaCie external SCSI Ultra320 units, on different external ports of the 39320. The Ultrium 460 tape drive is connected to the Windows side. The CDROM drive (a Yamaha SCSI 44x) is on the UNIX side).
HPTapeDiagnostics indicate no problem.

Previously, when I was using the Adaptec 3940AUWD, I was getting 307.69 MB/min using NTBackup. So, this is very perplexing. I'm supposed to be getting 4x the performance with the 39320.

[David Ruska]

Here's a document on performance tuning for Ultrium1 - everything appies to ultrium2 and then some:
http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=lpg50049

Using windows native backup is not recommended - it's not performanced tuned. I'm not sure what blocksize the native backup uses, but it may not be optimal (64K or greater recommended).

Run the HPTapePerf (or the integrated version in HP LTT) on the drive and see what performance you get. That will tell you if the problem is windows native backup, or the HBA/driver combination.

[Ronald W. Satz]

"Check with your ISVs if your application can
Increase tape block size
Increase memory allocation
Increase SCSI transfer size above 64K"

David:

Thanks, how do I do the above? As I said, the HPTapePerformance Program shows an acceptable speed. I do have many small files; however, when I was using the AHA-3940 (in the same system) I got much better speed.

Regards,
RWS
transpower@aol.com

[David Ruska]

> "Check with your ISVs ...

> Thanks, how do I do the above?

If you're using windows native backup, I don't think you can tweak any of the settings mentioned. You could try the trial version of tapeware (www.tapeware.com). Other supported ISV apps can be found here (for XP):
http://www.hp.com/products1/storage/compatibility/tapebackup/software/detail/17-0099-0136.html

> As I said, the HPTapePerformance Program shows an acceptable speed.

I wasn't sure if you had run that under windows XP or under Unix.

> I do have many small files; however, when I was using the AHA-3940 (in the same system) I got much better speed.

You can run the "sys perf" tests in HP LTT to get a feeling for how fast you can get data off your system. Small files will definately slow things down.

[Ronald W. Satz]

I tried connecting the Ultrium 460 to the A Channel (disconnecting the UNIX drive). This way, Windowx XP is on the B Channel and so the backup is going across the channels of the 39320. The tape drive is self-terminated; the Windows drive is actively terminated. Alas, the backup time using NTBackup turned out about the same (51 minutes for 7.9 GB). I switched the cables back to the way they were and got, again, the same time. Using NovaBACK helped a little bit, doing the backup in 43 minutes and 50 seconds, for a speed of 186 MB/min. But this is still about 10% of the rated speed and much less than what I got on the 3940.

I can try Tapeware. Is there any other software I should try which is known to be much faster than NTBackup?

[Dave Dewar]

Hi Ronald,

Could you run the "sys perf" tests in HP LTT to get a feeling for how fast you can get data off your system as Dave Ruska suggests please.

I don't know why you are seeing different performance between the 3940 and 39320 HBA's but suspect it may well be driver related. However, I would really like to see what you system is capable of in terms of reads/writes to disk before we try to tackle the 3940/39320 performance differences.

The SysPerf tests in LTT will allow you to test "backups" of your filesystem to memory and "restores" of memory to your filesystem and should give provide you will an idea of how quickly your system(disk,processor,memory) can move data to/from disk.

The "lowly" pentium 2, single 7.2k rpm scsi disk, system I have on my desk for trying out LTT and other tools with our tape drives achieves 27-30MB/s on LTT DevPerf (HPtapeperf) tests with an Ultrium460, whereas you are only seeing 19MB/s. When running Sysperf backup pre-test on this system I can get anywhere between less than 1MB/s and over 20MB/s depending on block size and the files I ask to test to use.

Cheers,

Dave Dewar

[Ronald W. Satz]

Dave:

The SysPerf test gives a result of 3 MB/sec, with version 3 of the Adaptec Ultra320 driver. When I was using version 2 of the Adaptec Ultra320 driver, I was getting only 2 MB/sec. The LaCie drive is rated at 68 MB/sec. The PassMark Performance Test v. 5 gives a Disk Sequential Read of 27.2 MB/sec and a Disk Sequential Write of 13.9 MB/sec. I've tried different block sizes in NovaBack, ranging from 512 to 4096--1024 (the default) is marginally the best. I've orderded the Tapeware trial CD and will try that.

Regards,
RWS
transpower@aol.com

[Dave Dewar]

Hi Ronald,

Those Sysperf results are very low. You mentioned earlier in the thread you were getting about 186MB/min backup transfer rate with Novaback. Well the 3MB/sec you are getting with Sysperf would translate inot 180MB/min transfer rate.

I pulled the specs on the LaCie and it is a 10K rpm disk with a headline spec of 68MB/sec as you say. However, disk drive specs are typically the worst of the worst in terms of "what you see is what you get". What the spec fails to mention is that this 68MB/sec is only available if you are on the outer diameter of the disk and are moving along a continous track. The inner diameter transfer rate is typically half this rate since the number of sectors per track (data per track) is half. So now you have a spec that goes from 30-68MB/s. Add in a small amount of fragmentation and small files and this will go even further down as the head is moving across the disk from track to track reading each dispersed file and you start to take hits due to track to track positioning delays of about 0.5ms and track seeks of about 4ms.

Your passmark tests are sequential tests, i.e. reading continously along a track, hence the values you are getting from these are probably your true maximums. However, for Sysperf to be so much less than this does suggest you must have a heavily fragmented disk, or at least in the areas you asked Sysperf to backup, the formatted block size of the hard disk and the size and number of files you are backing up are a factor as well.

I don't think going to Tapeware is going to help you much until you get the performance of your disk transfers up. The performance trouble shooting guide that Dave listed above does provide some help here. Basically, you backup/restore rates appear to be limited by your disk transfer rate There is some other factor due to different ISV apps and SCSI HBA's that appears to be occurring as well, but the disk transfer rate issue is the dominating factor.

Cheers,

Dave Dewar

[Ronald W. Satz]

Dave:

I defragmented the LaCie before the last set of backups, so we can rule out fragmentation as a problem.

Could there be something in the SCSI configuration that's a problem?

Regards,
RWS
transpower@aol.com

[Dave Dewar]

Hi,

If if was a SCSI config issue I would expect to see problems during your passmark results and transfers from memory to tape earlier as well.

If you have defragmented the disk we can rule out fragmentation. However, if you have a lot of small files as you mention you can get a similar effect since the disk is spending a high proportion of time locating the next file compared to actually transfering data,

a good link from the performance trouble shooting guide Dave posted, I have included it below explains this well.

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=lpg50032&locale=en_US

What I would like you to do is follow the link above and download PAT as it suggests. The link to download PAT is given below.

http://h20000.www2.hp.com/bizsupport/TechSupport/SoftwareDescription.jsp?locale=en_US〈=English+%28US%29&pnameOID=12188&prodSeriesId=42846&prodTypeId=12169&basePartNum=co557&locBasepartNum=co557en&los=Microsoft+Windows+2000&tech=Software

PAT is a small standalone performance analysis tool for disks that allow you to measure both file system access and sequential access to the disk. The SysPerf tool in LTT only allows file system access.

Run PAT in both modes. Running in sequential mode should bypass any "small file, defragmentation" issues and you should get similar performance to your passmark results, the file system mode should give you the 2-3MB/sec result you saw in LTT Sysperf.

If you want me to check you scsi configuration at the same time, you can capture a LTT support ticket, save in ltt format, zip up and attach here and i'll see what ltt says about your system and scsi config.

Cheers,

Dave Dewar