Re: Informix on Raid 5

Murray Oldfield wrote: > > Some RAID Array implementations such as Data General (or OEM) CLARiiON > allow large cache (64MB) mirrored across two storage processors with > battery backup to allow flush to disk on power failure. Use of a Global Hot > Spare and throttle for rebuild will cut down the hit when a disk does fail. > As rule RAID 1/0 with more spindles will be fastest. > MO > > Hutch Hutchison <hutch@gnorth.com> wrote in article > <32F8A9E6.2B19@gnorth.com>... > > > My programmers tell > > > >me that the system would be faster useing a multidrive config. than > with > > > >Raid. Please respond with any experiences or knowledge regarding this > > > >subject. > > > > > > > > > > I believe RAID 5 is the version which has very poor write > performance. > > > This is because all of the parity info is stored on one disk. This > > > means that EVERY write involves one write to the parity disk. > > > > > > You would get better peformance with RAID 1 (i.e. mirroring) where > > > not all writes involve writing to the same disk drive which becomes > > > a bottleneck. > > >Actually, the performance hit is not because it must read the parity disk. In RAID 5, parity is spread across every disk. RAID 3 has a dedicated parity disk for a rank of data disks. In RAID 5, the write performance problem comes from having to read parity, calculate new parity, then write data+parity. So every write becomes a read-write, which results in a performance penaly of 10-40% depending on whom you believe. Many RAID vendors, like EMC or Storage Computer, have implemented things like large caches and dedicated processors to lessen the hit of RAID 5, but you will pay more $/GB. Also, check your DB statistics. In our case, most of our tables had a 100:1 ratio of reads to writes. In this case, who cares about writing? We saw a large performance *increase* when we installed an NPI RAID5 subsystem in place of our non-redundant SCSI disk array (12 drives on 3 controllers). This was because the RAID array allowed more parallelism in disk reads. For a good article on this subject go to: http://www.compumedia.com/~kkirk/raid.html -- Matthew C. Reprogle Delco Electronics - IC Integrated Manufacturing phone: (765)451-9651 Internet: mcreprog@mail.delcoelect.com FAX: (765)451-8230