Spindles 1

[KenCunningham]

Hi folks. One of our suppliers is insisting (so far) that we make 20 spindles available for an Oracle app they are installing (mainly transactional use). This will add significantly to the cos of the development (I was proposing a twelve-spindle Sun Fire V880). I have argued, so far unsuccessfully,that disk speeds are now such that the spindles issue is redundant. I'm quite prepared to be shot down in flames here if any one can give me a case for spindles which is coherent and logical - so far my suppliers have been unable to do that. All views welcome. Regards.

 

[flstffatboy]

Ken,

Some of the biggest performance gains I've seen have come from increasing the amount of spindles especially with a transaction heavy Oracle App.

For your new system what type of RAID do you plan to use? The read write activity associated with the type of RAID you plan to use is very important to understand if you should use more disks.

FLSTF

 

[alexhu]

Ken

Take it to the extreme - would you put the database on one big disk - why not? What made you pick on 12? Why not 10 or 8? The more you can spread the I/O the better the performance of the disk subsystem.

Alex

 

[KenCunningham]

My main 'problem' if that it is, is that I was proposing to use a fully stacked Sun Fire V880 (12 73 Gb disks), in a RAID 5 configuration (excepting the system files), but the configuration proposed by my suppliers means that we will have to spend an extra £20,000 or so on a Storedge cabinet and disks. I take flfstfatboy's point - can he give any indication of the degree of performance gain he's experienced, please? Again, many thanks for the points made so far. Cheers.

 

[DBAwhosaysNIE]

Ken,

I gonna whip out my flamethower at the other responses & agree w/you. if you're building a data warehouse where you'll be scanning TBs of data then by all means use as many spindles (& controllers which neither responder mentioned) as you can get your hands on. if you're building an app which does lots of small random reads/writes 20 vs. 12 won't make a hill of beans difference. FWIW, my instances do selects for ~20M page/day website while doing enough DML to switch a 100M logfile every ~15/min. storage? 7 drive (10K 160Mb/s) RAID 5 (w/32MB cache on controller) per box. how? propery tuned SQL, SGAs & storage clauses. if you want to introduct additional spindes for fault tolerance that's fine (assuming they're on separate controllers) but as performance argument it doesn't hold water for a system serving OLTP.

just my 2c but (based on what you've described) I'd say your consultants are FOS and I have a fair amount of experience with high-volume Oracle. one final thought, if you're trying to save $ why are you looking at Sun? Oracle runs incredibly well on Linux for a fraction of the cost (hardware anyway).

 

[jimbopalmer]

My most insane Oracle instance was in 1994 and used 130 1 gig drives 120 in a mirrored striped configuration, 4 system disks and 6 application drives. it was very fast but every 6 months when I powered down the 4 arrays I had at least one drive in failure.

(BTW, Striping Mirrors is better than mirroring stripes!)

I tried to remain child-like, all I acheived was childish.

 

[KenCunningham]

DBAwhosaysNIE, Jimbo,

again many thanks for your responses. I must admit that for OLTP applications such as the one we're planning, DBAwsN seems to sum up my thoughts on the matter. I'll bring it up at the next project team meeting and ask the suppliers for a considered and technical reply. I agree Jimbo, the more hardware you have, the more to go wrong! All the best and thanks to all again.

 

[mhyman]

I have to disagree.

The more spindles you have is important, but RAID level is critical to Oracle. Using RAID 5 is a huge potential mistake as write performance is awful.

If you have a mostly read -mostly environment like a data warehouse, its fine, but when you need to write fast, stay away from RAID 5.

So, if you accept that and then still want redundancy, you should use mirrored RAID 0, or stripes. Is that right?? Well what I mean is to make stripes across many spindles and then mirror that array over a different controller.

You mentioned that cost was an issue, but I would recommend 2 storage arrays, with one mirroring the other. That way, even if an entire device fails, you have a duplicate still online. If your needs are that tough, than a standalone RAID system is okay.

Optimizing stripe size and memory acting as a buffer or cache will go very far, with Oracle in adding performance.

This is of course, just my opinion based on a lot of experience.

Good luck.

 

[DBAwhosaysNIE]

how recent is your experience that write performance on RAID 5 is awful? 2 yrs ago I benchmarked a 12 drive Sun A1000 (attached to E4500; prod @ time) 0+1 against a 7 drive RAID 5 in an IBM EXP300 (to x350; eval @ time) with an off the shelf RAID card. test was restoring hot backups of ~60GB databases (not exhaustive but most write intensive thing I do) and the eval system consistently smoked the old one. why? 160 vs 40 Mbs drives for one but also controller had 64MB cache (battery backed) and 200MHz processor handling parity (the infamous RAID 5 &quot;write penalty&quot. my prod instances have already generated >5GB of archive logs since midnight which is nothing by warehouse standards but way more than any ERP or CRM system I've ever worked on (which is quite a few). 0+1 may still be faster on identical hardware but for <$5K you can get a storage solution that will roast one costing 10x as much at the height of .com days.