RAID 5 versus mirroring

Ok guys, I've seen several posts regarding RAID 5 versus mirrored arrays. Maybe a disk guru can help me out here...I'm merely a lowly programmer. When running vmstat on our production dbserver, an RS6000, 6GB mem, 12 CPU's @ 262 Mhz with a software-mirrored array, 16 disks x 2, plus 2 for the OS, we're seeing the I/O wait part of the vmstat output in the 60-90% range. When we take a Level 0 backup and apply to another RS6000 with 2GB mem, and 4 340 Mhz CPU's, and 10 disks in a RAID 5 arrangement, we see vmstat showing 0-1% I/O wait. All machines are running AIX 4.3.1 with Informix 7.3.1 (which I think desperately needs to be upgraded, right?) So the question is - what exactly does the IO Wait part of vmstat mean, and is it indicative of poor disk I/O? Also intersting to note - running a single 4gl program on either box yields very similar performance results. Our db activity is mostly reads versus writes judging by looking at sysmaster information - bufreads, bufwrites, isam reads,writes, deletes - etc. When I sum everything up, reads outweigh writes by about 1000 to 1 or more on every table in the system. Hopefully this is a reasonalby accurate way to figure this? But I understand the read/write ratio is I've taken a look at how the DBA laid out tables on the mirrored machine, and only the largest tables (3 millions rows +) have fragments in multiple disks. For all the rest, the tables reside entirely on one physical disk. I suspect that a better layout would help out, but I don't know how much. Our management is asking for recommendations on ways to get more speed out of our machines. We can do whatever we want - hardware, software - whatever. But we need to spend money on the right stuff, of course. Any thoughts you can provide would be most appreciated - I've read lots of good info out here. Thanks!

NO RAID5 NO RAID5 NO RAID5 NO RAID5 NO RAID5 NO RAID5 NO RAID5 That said, one thing you do not say, which is apparent from the results is that the mirroring is being done in software, whether the OS mirror drivers or in Informix, which I can see because of the high CPU cost of the mirroring while you RAID5 implementation is obviously controller or server based. (Note that the CPU cost of a software RAID5 solution is even higher that mirroring.) To your last question if you trade the RAID5 and mirrored singleton disks for a hardware RAID10 solution, such as EMC's or DG/Clariion's, you will be able to spread I/O's across all 16 drive pairs instead of just balancing each I/O over the one mirrored pair. This will improve read performance over BOTH the mirrors and RAID5 while increasing your protection over RAID5 immensely. Write performance will improve, but, as you point out, in your case write performance is not the problem. Note that my main objection to RAID5 is NOT performance but safe of you data. Obnoxio recently posted my quarterly RAID RANT. Art S. Kagel Eric Fontaine wrote: > > Ok guys, I've seen several posts regarding RAID 5 versus mirrored arrays. > Maybe a disk guru can help me out here...I'm merely a lowly programmer. > > When running vmstat on our production dbserver, an RS6000, 6GB mem, 12 CPU's > @ 262 Mhz with a software-mirrored array, 16 disks x 2, plus 2 for the OS, > we're seeing the I/O wait part of the vmstat output in the 60-90% range. > > When we take a Level 0 backup and apply to another RS6000 with 2GB mem, and > 4 340 Mhz CPU's, and 10 disks in a RAID 5 arrangement, we see vmstat showing > 0-1% I/O wait. > > All machines are running AIX 4.3.1 with Informix 7.3.1 (which I think > desperately needs to be upgraded, right?) > > So the question is - what exactly does the IO Wait part of vmstat mean, and > is it indicative of poor disk I/O? Also intersting to note - running a > single 4gl program on either box yields very similar performance results. > > Our db activity is mostly reads versus writes judging by looking at > sysmaster information - bufreads, bufwrites, isam reads,writes, deletes - > etc. When I sum everything up, reads outweigh writes by about 1000 to 1 or > more on every table in the system. Hopefully this is a reasonalby accurate > way to figure this? But I understand the read/write ratio is > > I've taken a look at how the DBA laid out tables on the mirrored machine, > and only the largest tables (3 millions rows +) have fragments in multiple > disks. For all the rest, the tables reside entirely on one physical disk. > > I suspect that a better layout would help out, but I don't know how much. > Our management is asking for recommendations on ways to get more speed out > of our machines. We can do whatever we want - hardware, software - > whatever. But we need to spend money on the right stuff, of course. > > Any thoughts you can provide would be most appreciated - I've read lots of > good info out here. > > Thanks!