Having looked at this case, it seems like the issues are settled. +1 - Garrett
On Wed, 2010-07-28 at 07:17 +0800, bz211116 wrote: > Nicolas Williams wrote: > > On Wed, Jul 28, 2010 at 07:06:20AM +0800, bz211116 wrote: > > > >> Nicolas Williams wrote: > >> > >>> On Tue, Jul 27, 2010 at 12:47:31PM -0700, Garrett D'Amore wrote: > >>> > >>>> Is there a reason that this we would ever, under normal circumstances, > >>>> want to use RMW on these devices? Is there a reason that this should > >>>> even be exposed as a tunable to customers? > >>>> > >>> Probably when accessing UFS, FAT, and other such filesystems. > >>> > >> Yes, For other file systems or applications which still issue non-4K > >> aligned I/Os to those SSDs which has low performance RMW in f/w, turn > >> on RMW in sd can greatly improve the performance, our experiments show > >> 100x faster than RMW in f/w. But for those disk drives which perform > >> RMW better at f/w, we should turn RMW in sd driver off. This is one > >> reason we need this tunable. > >> > > > > But do we care about performance of non-ZFS here? (Answer: probably. > > I'm thinking of raw device uses too.) > > > > > >>> But I don't get why this needs to be a tunable either, at least for ZFS, > >>> since we could ensure that ZFS always does 4KB aligned writes, and then > >>> who cares if UFS, FAT, and friends run slow on flash. > >>> > >> For ZFS, which now uses READ CAPACITY 16 (if succeed) to get the > >> physical block size of SSD. If the physical block size is 4096, ZFS > >> aligns its minimal I/O size and request address at 4K boundary which > >> can get best performance. But most of the SSDs now still report 512B > >> physical sector size or even do not support physical sector size at > >> all, and some of them also has bad RMW performance in f/w. in this > >> case, we should turn on RMW in sd for them. that's another reason we > >> need this tunable. > >> > > > > But shouldn't ZFS just always do 4KB aligned writes and be done? Who > > cares if the SSD claims to have a 512B physical sector size? > > > Most of SSDs in F5100/F20 is in this case... > > Thanks. > -bo > > Nico > > > _______________________________________________ opensolaris-arc mailing list opensolaris-arc@opensolaris.org