D2D backup solutions for VMS

[robert70]

Hello,
We ahve a DS20E Alpha Server running VMS 7.3-2
and hosting an Oracle Database 9i on a StorageWorks MSA1000 RAID Array.

We are finding backup and Restore operations onto SDLT Tape taking too long. Our Restore Operation of the Database takes about 3-4 hours - 175Gb of data compressed

Can you advise on possible Disk to DIsk solutions to enhance our backup eg keep the magnetic tape backup say once a week but do a D2D daily backup -

Thanks

[Robert Gezelter · Accepted solution]

Robert,

Does "backup" mean an image of the entire volume, or does it mean a backup of the data in the database.

If one means backup of the drive, then the least downtime is generally adding an additional member to the shadow set, waiting for it to become synchronized, quiesce the volume (ensure that the database is consistent), disconnecting the spare shadowset member, and then continuing operation.

This does require that the database and the applications cooperate.

The backup volume can be retained online, or backed up to tape while the system continues operation. If the original volumes need restoration, the backup disk can be used as the source for a replacement shadow set (and a base for applying audit journals).

- Bob Gezelter, http://www.rlgsc.com

[robert70]

It would be to backup the data and not the volume - we currently have the RAID configured with NO shadow sets but but using RAID 5 - striping etc for the database backup.

how would it work in this case?

[marsh_1]

HI,

is the tape drive a standalone or part of a library, is it direct attached or san attached, the msa1000 again is that direct attached or san attached. do you have access to a san with other storage resources. are there any current backup software solutions employed at your site that you could utilise .

hth

[robert70]

Tape Drive stand-alone connected via SCSI
No other SAN available in office.No other resources available.

[robert70]

The MSA1000 is directly connect to the Alpha via fibre optic cable.

[marsh_1]

hi,

so it sounds like you have a substantial backup window (potentially 3-4 hours) where the database is down.
do you have room in the msa1000 for more disks, or budget for another shelf/disks ?
if so create a new raid5 or 1+0 lun for disk to disk backup (suggest raid 1+0 for better write performance if you can afford it) once complete you can get a tape backup done from the backup volume while the application and database is brought back up. roughly speaking you will get about up to 150 mb/s max reading from the raid 5 , writing to the raid 1+0 at 100 mb/s max.
back the other way you are looking at 150 mb/s read from the raid 1+0 , 68 mb/s write to the raid 5 for restore.
i'll leave you to do the maths but it should substantially reduce your backup window, a daily tape backup will provide extra redundancy with only impact being the system resources required for a tape backup.
the other way as bob has suggested still involves creating extra luns and implementing shadowing.

hth

note:- above figs based on 64k sequential reads/writes with 14 drives (36gb 15k rpm), different configurations (ie slower or larger drives) will affect performance.

[robert70]

thanks
yes we have room for another 10 disk arrays in our config so we could place in a few 36.4 gb drives.
Would the backup on D2D be standard or do you normally compress?

[robert70]

it take 1+0 setup is a shadow set config?
we have raid5 eg one stripping disk in 2 disk volumes - dkb200 and dkb300 both have 145gb physical space - 5 x 36.4 Gb in each

[Robert Gezelter]

Robert,

RAID 0+1 has far higher performance than RAID 5, particularly for writes (as disk prices have come dramatically down, the balance has tilted towards RAID 0+1 in terms of price performance).

Using Host Based Volume Shadowing to bring an extra set of drives into the shadow set allows the downtime window to be reduced to mere moments, as I noted in my post. The duplication of the data occurs in parallel with normal operation. The extra volumes can then be remounted privately for tape backup.

Generally, what is needed is the licensing for host based volume shadowing, and the extra drives.

Note that this is different from controller based RAID.

- Bob Gezelter, http://www.rlgsc.com

[marsh_1]

hi,

raid 1+0 would be implemented in hardware in the msa1000.

hth

[robert70]

mark
i think we have 128k file sizes (minimum 273 blocks on our small files on database drives)would this improve or slow down the resore process.

Am i correct in assuming our current 185Gb of data could be restored on raid5 is 58mbs eg take about 45 minutes?

[Hoff]

The usual approach for trimming your window uses host-based volume shadowing; host-based RAID-1) as Bob is suggesting. Possibly in conjunction with what Mark is referring to with the controller-level RAID-10, too.

Details on splitting off volumes are in the recent OpenVMS volume shadowing manual. In general terms, you quiesce the database operations, split off the volume, and off you go. You have a disk with a parallel copy of the database and the rest of the data on the disk, and can spin it off to tape at your leisure.

The downside of this technique applied to a database is the possibility of inconsistent data due to in-flight changes. Which is why most production databases including Oracle offer archival processing. Similarly, with MySQL, you'd use the mysqldump tool, and with Rdb you'd use RMU /BACKUP.

SDLT and Ultrium tapes are often faster than disks for various common operations, so I'd have to wonder if the host can supply the data fast enough to keep the drive streaming. SDLT and Ultrium are very sensitive to speeds and feeds, but (when they're streaming) very fast.

BACKUP on OpenVMS is sensitive to the process quotas and (when quotas are tuned correctly, and dependent on the particular setting of /LOAD on recent versions of BACKUP on OpenVMS) can consume an entire processor; it'll drive a tape drive at pretty close to its theoretical limit. The usual limits are then I/O bandwidth, I/O contention or (commonly) disk fragmentation.

http://labs.hoffmanlabs.com/node/49

I'd expect Oracle also has quota recommendations.

[marsh_1]

hi,

which one you pick depends on your needs :-

disk to disk backup - requires an application outage , just purchase disks so is the cheaper option and simplest to implement

shadowing - virtually no application outage, requires disks and software license so is the more expensive and complex option.

fwiw

[Hoff]

> Am i correct in assuming our current 185Gb of data could be restored on raid5 is 58mbs eg take about 45 minutes?

Consider fully testing your recovery path. That testing will also and entirely incidentally tell you how long the process requires. And more critical than the elapsed time required, whether the particular recovery produces a functional result.

[marsh_1]

hi,

those figures are test lab results, so they are possible but as hoff has stated the only way to find out your 'real-world' figures is to test the solution fully.

fwiw

[robert70]

many thanks

[Hoff]

nb: RAID-5 engenders massive I/O loading during spindle recovery, and is also surprisingly often clobbered by double-spindle failures; empirical data shows the spindle failure rates spike to 4x normal during a RAID-5 recovery.

http://labs.hoffmanlabs.com/node/410
http://labs.hoffmanlabs.com/node/93

Having looked at the costs of the disk spindles and at the data around error rates, it's been years since I've configured RAID-5 anywhere.