Re: Informix disk mirroring

In article <3ueg3n$5qg@host-04.colby.edu>, dwcooley@colby.edu (David W. Cooley) wrote: > We are running INFORMIX/Online 5.0 with INFORMIX based mirroring on an > HP 9000 G40 running HP-UX 9.04. We're running 4.11 on Pyramid SMP hardware, and their DC/OSx 1.1 operating system (essentially SysVr4). We have been running various versions of Informix on various platforms for about 6 years. > When we initially installed the software we discussed the tradeoffs of > HP-UX LVM based mirroring and the INFORMIX database engine based > mirroring. We explored this issue with our application vendor and > with INFORMIX. Both stated unequivically that we should use the > INFORMIX based mirroring because the database engine has a much better > understanding of the data than the operating system does and thus > would optimize i/o... For what it's worth, we have =always= used the Pyramid "Virtual Disk" facility, which provides mirroring at a level between the device driver and the rest of the kernel in a manner that is essentially invisible to both the kernel and any applications. We do so primarily because the tools to manipulate the mirrors are so much more robust, as is the system's tracking of the state of the mirrors across failures, panics and reboots. It also provides n-way mirroring if desired, including across dissimilar devices (such as including a RAM drive as a mirror piece). Further, the Virtual Disk offered much better performance than Informix mirroring. Informix (at least in the versions we last tested) always reads from the first piece that is on-line. Pyramid reads from the piece with the shortest queue. The result is reads are substantially faster, especially on a loaded system (or if a RAM drive is included as a mirror piece) There are other performance advantages O/S mirroring can offer. For example, when there is a queue of work for a particular disk spindle, Pyramid (and others) maintain that queue with an "elevator sort" so the I/O will be done in sequential cylinder order, substanially reducing head motion. Informix cannot do that, lacking knowledge of the layout of the various chunks that make up a dbspace. We benchmarked our software on such platforms as IBM RS/6000 hardware running AIX--in that case we used the LVM mirroring. On the Tandem S2 we also used the O/S mirroring facility. The only time we've allowed Informix to provide the mirroring was on a very low-end (PC platform) implementation where no other choice was available. > ...and would processes transactions correctly (there was a strong > implication that o/s based mirroring would lead to data corruption). > So we went with the advice but skeptically. We have never seen data corruption that could be traced to a failure of the O/S to maintain mirror integrity. > Do others have any experiences or information on this issue which might > shed some light on the matter? Does that help? > -- > David W. Cooley e-mail: dwcooley@colby.edu > Associate Director for Technical Services phone: (207)872-3581 > Computer Services fax: (207)872-3555 > Colby College > Waterville, Maine 04901 -crl -- Chad R. Larson (CRL22) Brother, can you paradigm? 602-870-3330 HRFR12A@prodigy.com ...!anasaz!chad chad@anasazi.com Anasazi, Inc. - 7500 North Dreamy Draw Drive, Suite 120, Phoenix, Az 85020