Re: EMC vs. ordinary RAID 5

---------------------- $BE>Aw<T(B: michael martin/G/KMT/JP $BE>AwF|(B: 2000/03/30 17:55 --------------------------- $BAw?.<T(B: michael martin $BF|IU(B: 2000/03/30 17:25 $B08@h(B: Stefan Weideneder <stefan@weideneder.de> cc: $B7oL>(B: Re: EMC vs. ordinary RAID 5 (Document link not converted) Here is one of Art S. Kagel's explanations of why not to use RAID5 (with apologies and thanks to Art): "There are two problems with RAID5. <SNIP> since the one pair comprises only 20% of accesses, performance is affected no more than 16% during recovery and the risk of catastrophic data loss is reduced." I have posted this in full to Stefan. Sorry I can't remember the URL where I found this but if anyone wants the full text I will be happy to mail it to you if that's OK with Art ? Michael Martin michael_martin@komatsu.co.jp

Hi Michael, the explanation is really good. But all the information regarding the performance is somehow needless. It would be interesting, if someone could process much more than 65MB/sec from 6 disks. It's up to you to run a small test on your own machine while you run a PDQ query or a parallel backup with ON-Bar or ON-Archive. Simply look at the blocks that you can read from your disks. If your sar-output shows more than 130000 blocks ( sum of all disks ) during these operations, then you can say that your system is faster than the 20,000$ DELL machine. A point of interest for most customers I think is the price that they have to pay for their systems. Since customers who need a high availability system will use two boxes and DR, I don't believe that the complete data loss is that important. Finally, Art wrote a lot of good articles in this NG. But Art is not responsible for your performance. Run the test above if you followed his instructions. Best regards, -- Stefan Weideneder Phone: +49 89/3565478-2 --------------- --- Fax: +49 89/3565478-3 ------------- ------ mailto:/stefan@weideneder.de --- -------- http://www.weideneder.de -----

Stefan Weideneder wrote: > Hi Michael, > [SNIP] > > A point of interest for most customers I > think is the price that they have to pay > for their systems. > True. But what is peace of mind worth and how much does it cost to drag your DBA out of bed several times a quarter? How much is it worth to know that none of the users on either of your servers is getting garbage instead of data? > Since customers who need a high availability > system will use two boxes and DR, I don't > believe that the complete data loss is that > important. > I run 6 servers for our news service and 4 other server pairs, I still insist that ALL of our disks for Informix are configured as RAID10. (Truthfully I don't give a S**T where the sysadmin places the TEMPDBS chunks, but the rest... ;-) > > Finally, Art wrote a lot of good articles in > this NG. But Art is not responsible for your > performance. Run the test above if you followed > his instructions. I agree whole heartedly, and if performance were the only reason to oppose RAID5 I'd have shut up 3 years ago and been happy to feel smarted that the average bear. Certainly there are many installations where the extra performance is not an issue and should not be part of the equation. However, the real issue is the HORRENDOUS problem of RAID5 not verifying checksums on reads and therefore not detecting partial media failure. If you must save a buck (or sheckle or what-ever) use RAID3 or RAID4 instead of RAID10. All of the arguments for RAID5 over these configurations are based on random read performance and I submit that Informix engines do not ever do any I/O that looks like the small block random I/O that RAID5 proponents trot out. At least any decently implemented RAID3 or RAID4 implementation checks things on read. Art S. Kagel