On Tue, Aug 26, 2008 at 06:36:42PM -0400, Thor Lancelot Simon wrote: > On Tue, Aug 26, 2008 at 02:30:20PM -0700, Christopher Chen wrote: > > On Tue, Aug 26, 2008 at 1:45 PM, Aaron J. Grier > > <agrier%poofygoof.com@localhost> wrote: > > > On Mon, Aug 25, 2008 at 11:47:18AM -0400, Thor Lancelot Simon wrote: > > >> The real fix for the "writes metadata twice" issue is to support a > > >> journal on another partition. With a real SSD (that is, one that's > > >> not built from flash memory with its insane write penalty) this would > > >> make the extra writes almost free. > > > > > > prestoserve! > > > > YAY! > > > > Do you actually have one? > > There's at least one vendor ("vmetro") of very large PCI NVRAM cards, > [...] > These cards are *not* cheap. Solid-state disks with SAS or SATA interfaces > are likely to be a lot more cost-effective. These are the ones I'm interested in: http://www.hyperdrive4.com From a brochureware point of view they seem very promising, though I've never laid hands on one. I was looking at them primarily for postgresql WAL and temp tablespace use, with a secondary interest in using them for detached ZFS ZIL. For the pure write log use case (whether db, zfs, wapbl, or other filesystem) they're quite large in cost-effective configs, so it's nice to have something else that can usefully take advantage of the rest of the space. > What would really rock would > be firmware for one of the NVRAM-equipped RAID controllers that allowed > small volumes to be created which existed only in the controller NVRAM. That's the model being hoped for in some Sun arrays, because it plays very nicely for ZFS detached ZIL. -- Dan.
Description: PGP signature