On Wed, Jun 30, 2010 at 01:55:30PM -0500, Marco Peereboom wrote: > New drives have much more sophisticated IO queueing algorithms and > posses much more information about the physical limitations of drive > platters (or lack thereof) than an OS can ever hope to have. So switch > SCSI (and SATA) drives to use the shiny new bufq that does not sort IOs. > Essentially this keeps IOs closer to the drive. > > The impact I have seen is negligible noise in sequential IO however > random IO seems to get a 20 - 25 % performance benefit. > > I'd like to commit this soon so speak now or forever hold your piece. >
ok krw@ .... Ken > Index: sd.c > =================================================================== > RCS file: /cvs/src/sys/scsi/sd.c,v > retrieving revision 1.198 > diff -u -p -r1.198 sd.c > --- sd.c 28 Jun 2010 08:35:46 -0000 1.198 > +++ sd.c 30 Jun 2010 18:38:20 -0000 > @@ -182,7 +182,11 @@ sdattach(struct device *parent, struct d > */ > sc->sc_dk.dk_driver = &sddkdriver; > sc->sc_dk.dk_name = sc->sc_dev.dv_xname; > - sc->sc_bufq = bufq_init(BUFQ_DEFAULT); > + > + if (SCSISPC(sc_link->inqdata.version) >= 2) > + sc->sc_bufq = bufq_init(BUFQ_FIFO); > + else > + sc->sc_bufq = bufq_init(BUFQ_DEFAULT); > > if ((sc_link->flags & SDEV_ATAPI) && (sc_link->flags & SDEV_REMOVABLE)) > sc_link->quirks |= SDEV_NOSYNCCACHE;
