RAID Technolgy
Posted in 1999
Topics: General Discussion
We are using Informix Online Dynamic Server 7.2. We are going in for a new DEC UNIX Server with RAID Technolgy. It incorporates mirroring at the hardware level. So is it not necessary to do mirroring using Informix?. Can anybody give suggestions regarding mirroring and other features if your organisation has got a similar setup with Informix server & RAID? Sampath
The Informix "rule of thumb" is don't use RAID. And some of their reasoning does make sense. The engine has been optimized to make use of mirrors if they exist and will go so far as to split queries to make use of parallel processing. However, mirrors are a (relatively) costly hardware investment in both disk and CPU and require planning out one's disk farm. In my experience, what Informix 'wants to see' and what the people who make procurement decisions choose are sometimes at odds. When one's entire disk farm consists of 5, 9 gbyte drives in a Clarion array (because that was the most bang for the buck) and you have to be able to provide for recovery and (by the way) the application needs a little over 3 gbytes then RAID 5 provides a tradeoff. You let the RAID hardware do it's thing and you hope (!) that when one of those striped drives fail that the hot spare will work as advertised.
FProse wrote: > > The Informix "rule of thumb" is don't use RAID. And some of their reasoning does make sense. The engine has been > optimized to make use of mirrors if they exist and will go so far as to split queries to make use of parallel > processing. However, mirrors are a (relatively) costly hardware investment in both disk and CPU and require planning > out one's disk farm. In my experience, what Informix 'wants to see' and what the people who make procurement decisions > choose are sometimes at odds. When one's entire disk farm consists of 5, 9 gbyte drives in a Clarion array (because > that was the most bang for the buck) and you have to be able to provide for recovery and (by the way) the application > needs a little over 3 gbytes then RAID 5 provides a tradeoff. You let the RAID hardware do it's thing and you hope (!) > that when one of those striped drives fail that the hot spare will work as advertised. NO NO NO I use DG Clariion arrays and I HATE RAID 5 on them even though DG puts 8 bytes of error correction code on each disk block I still have partial media failure trashing databases occasionally. RAID 5 and databases DO NOT GO TOGETHER period! Those 5 drives cost your employer ~ $5,000.00 making that Clariion fully mirrored with the same capacity will cost an additional $3,000.00. Hell the taxi fares to truck me in to the office and back home and to feed me at indecent hours a few years ago when my news machines were crashing left and right before I put my foot down and got them mirrored came to more than that! Any penny pincher who does not see that data integrity and uptime are worth a lousy 3 Grand needs a new, pointy, haricut and you can tell him/her I said so. BTW - whenever a drive, controller or power supply fails on our Clariions with RAID5 (fully and automatically recoverable BTW) the driver issues a READ or WRITE error to Informix and the engine marks the chunk down, even though the RAID system should be supplying data using parity without error! This has NEVER happened on a RAID 10 system. Art S. Kagel
This is exactly what we've been saying for years. Basically you "*must* have a true hardware-based RAID3 architecture to avoid these media problems (not the RAID3 architecture currently defined where write buffering is used to minimize the ECC write tradeoff). Obviously, you then need a large read/write cache to counter the lower random read performance generally seen with parallel RAID. Then you need this cache to be protected, of course (ECC and mirrored with BBU). Incidentally Informix derived their 'anti-RAID' sentiments from a review carried out at the IFX technolgy center on Clariion RAID where they discovered that RAID0+1 outperformed Clariion's other configurations. When I took IFX to task over this they basically disowned the report saying that Clariion had funded the review - IFX only has the resources to test equipment if it is provided by the vendor! Seems a crying shame that the whole IFX industry is then stained by the performance of one specific RAID product. Like it or not, this report *IS* the defining document on RAID as far as any IFX salesperson is concerned. There was one significant other discovery in this report, and that IFX, and I guess all the other DBMS's, need to have their RAID stripe and DB I/O page sizes closely aligned on this type of RAID. If alignment is not carried out then 25% or more of performance is just lost - vanishes. If the alignment is impossible then the chunk size needs to be small. This may or may not then lead to compromises with the profile of the application. The bottom line is that some RAID systems are just too slow to handle the load that IFX places with its fragmentation approach. Faster RAID = smaller problem. Also note that fragmnentation is a sales and marketing bullet point over the other DBMS's and IFX don't appear to want to let it appear that the choice of hardware does actually make a difference. GB ----------------------------- NO NO NO I use DG Clariion arrays and I HATE RAID 5 on them even though DG puts 8 bytes of error correction code on each disk block I still have partial media failure trashing databases occasionally.
I agree with you both. RAID 5 and The Server Formerly Known As OnLine are a no-no. That is, if you are at all interested in performance. The two architectures are at odds. They are simply not made for one another. To take a trivial example: most RAID/ logical volume management systems cite as a selling point that you can add space to a volume at any time without bringing the system down. But it wouldn't be available to OnLine! You'd have to add an extra chunk, and you can do this any time you want anyway and OnLine doesn't care whether the new space is in the same volume. RAID 5 hurts write performance hard, unless you throw huge amounts of cache at it, which violates OnLine's fundamental recovery principle of the Guaranteed Write. With disk space now available for almost nothing (at least in the PC/ NT environment), I fail to see that the cost arguments in favour of RAID can be justified. In article <36B8D78E.1FA2@bloomberg.net>, kagel@bloomberg.net (Art S. Kagel) wrote: > FProse wrote: > > > > The Informix "rule of thumb" is don't use RAID. And some of their > > reasoning does make sense. The engine has been > > optimized to make use of mirrors if they exist and will go so far as > > to split queries to make use of parallel > > processing. However, mirrors are a (relatively) costly hardware > > investment in both disk and CPU and require planning > > out one's disk farm. In my experience, what Informix 'wants to see' > > and what the people who make procurement decisions > > choose are sometimes at odds. When one's entire disk farm consists > > of 5, 9 gbyte drives in a Clarion array (because > > that was the most bang for the buck) and you have to be able to > > provide for recovery and (by the way) the application > > needs a little over 3 gbytes then RAID 5 provides a tradeoff. You > > let the RAID hardware do it's thing and you hope (!) > > that when one of those striped drives fail that the hot spare will > > work as advertised. > > NO NO NO > > I use DG Clariion arrays and I HATE RAID 5 on them even though DG puts > 8 bytes of error correction code on each disk block I still have > partial media failure trashing databases occasionally. RAID 5 and > databases DO NOT GO TOGETHER period! Those 5 drives cost your employer > ~ $5,000.00 making that Clariion fully mirrored with the same capacity > will cost an additional $3,000.00. Hell the taxi fares to truck me in > to the office and back home and to feed me at indecent hours a few > years ago when my news machines were crashing left and right before I > put my foot down and got them mirrored came to more than that! Any > penny pincher who does not see that data integrity and uptime are worth > a lousy 3 Grand needs a new, pointy, haricut and you can tell him/her I > said so. BTW - whenever a drive, controller or power supply fails on > our Clariions with RAID5 (fully and automatically recoverable BTW) the > driver issues a READ or WRITE error to Informix and the engine marks > the chunk down, even though the RAID system should be supplying data > using parity without error! This has NEVER happened on a RAID 10 > system. > > Art S. Kagel >