Optimum RAID-5 "interlace-value" for Oracle DB?

So this evening I'm creating a six volume, RAID-5 array and by accident, hit <RETURN> before entering the "interlace-value". Given it is a database application, I would normally use something small, say 64KB. But the default is 512KB . . . <woops!> I could not stop initialization short of rebooting the system, Solaris-10. Even then, it might just restart initialization so I might as well let it run.

So the question is: What is the optimum, RAID-5, "interlace-value" for an Oracle DB?

This is primarily a data collection system with rather modest, less than 1 MB/sec., data rates. There are reports to generate but we don't like reports and even they are just serial access.

Thanks,
Bob Wilson

ps. Prius association? . . . I drive one to work.

 

Well we have some lessons learned:

• 16 KByte ~ 820 GB, 128 KByte ~ 808 GB :: some loss of space
• 126 GB backup ~ 40 hrs restore vs expected 24 hrs

We'll do a clean-up reboot and start the application to see how well it works in practice. So far, RAID-5 appears 'less fast' but fast enough. I'm not one to let perfect be the enemy of 'good enough.' But most of all, I'm happy:

• a single drive failure won't fail the system - we can keep working while scheduling and effecting a repair
• one massive filesystem with I/O spread across six disks - equalizes the I/O (and wear) so hopefully the whole will last longer than one drive being 'beat to death'

The disk drives rotate at 10,000 rpm and have 10 KB tracks. So we are looking at sectors that span 128 tracks. In theory, one might be able to size the sectors to reduce one cylinder, movement. However, we have no control over sector placement which means there will always be some cylinder movement. But there are some interesting geometry models:

1. decrease the sector size reduces the probability of a cylinder change during a read or write
2. increase the sector size to increase the probability of track-to-track switch for fastest read or write

Based upon the single cylinder delay and rotational delay, one can calculate what should be an optimum 'within disk' sector size. But the model also needs to assume a 1/2 rotational delay going from disk-to-disk. An engineering rule of thumb says the inter-disk delays should be equal to the intra-disk delays.

Regardless, I'm reasonably happy with the performance. It is not a 'barn burner' but it only has to be fast enough to meet our performance requirements. We've already had a failed disk down the system for nearly a month and I'm not going through that effort IF I can help it.

What is the alternative? Create a massive RAID-0 drive and just reload the software and reload the database base . . . which is what we had to do the last time. With RAID-5, we have to option to delay this multi-week effort.

Bob Wilson

 

Thanks,

Oracle is part of another application so we're using the parameters the vendor configured. I was disappointed to find the vendor had no practical RAID-5 recommendations. But I noticed the Oracle configuration file showed a block size of 8K.

I suspect we're in good shape but we'll do some metrics before making further changes. Sad to say, I suspect the first level, problem team may not have access to more detailed, engineering team data.

Bob Wilson