begin quoting David Brown as of Sun, Dec 16, 2007 at 02:19:44PM -0800: > On Sun, Dec 16, 2007 at 02:11:59PM -0800, SJS wrote: > > >A hardware RAID controller can write to each disk in parallel, computing > >the necessary checksums, etc., while a JBOD with a single controller > >will need to write to each disk separately (if the data is to span > >multiple disks). > > Software RAID can do this as well, if you're willing to put each drive on a > separate bus, and the OS and the rest of the architecture can handle it.
Yeah, but that extra hardware starts pushing you in to hardware RAID territory. The line is a bit fuzzy.. > You do have to compute checksums with one of the CPUs on the host. This > becomes more of a problem with more complex parity algorithsm, such as > RAID-6 uses. I think we probably have enough excess CPU processing power to mitigate the cost of computing a checksum. > >That's not saying that hardware RAID controllers all *do* that; I > >would not be at all suprised to find out that some don't. But they > >*can* (and arguably *should*). > > If _would_ surprise me if the cheap "RAID" boxes you can buy and drop > drives in and appear as either a USB device, or a network share do things > well. It would surprise me if they even worry about power-failure > robustness. A UPS should help with that. Except I was forced to do a 120 bounce last night on one of my machines (not a Linux machine) -- "sync" hung. Not "paused to write a lot of data", but "hung and never returned". I suspect a deadlock, as I also had a process that kill -9 wouldn't touch. So I bounced the machine. Yes, I was impatient. Three hours was long enough. -- I'm glad I had a journalling filesystem. Stewart Stremler -- [email protected] http://www.kernel-panic.org/cgi-bin/mailman/listinfo/kplug-list
