> On May 25, 2017, at 11:21 AM, Ken Merry <[email protected]> wrote: > >> >> On May 24, 2017, at 6:39 PM, Andrew Gabriel <[email protected] >> <mailto:[email protected]>> wrote: >> >> On 24/05/2017 22:10, Ken Merry wrote: >>> Is anyone working on SMR support for ZFS? >>> >>> I put support into FreeBSD for SMR drives from the block layer (GEOM) >>> through the SCSI layer (CAM): >>> >>> https://svnweb.freebsd.org/base?view=revision&revision=300207 >>> <https://svnweb.freebsd.org/base?view=revision&revision=300207> >>> >>> So far, no filesystems in the FreeBSD tree are using the SMR support. >>> >>> It looks like there are Linux folks working on SMR support: >>> >>> http://events.linuxfoundation.org/sites/events/files/slides/lemoal-Linux-SMR-vault-2017.pdf >>> >>> We (Spectra Logic) are considering starting work on supporting Host Aware >>> and Host Managed SMR drives in ZFS, and we’d rather collaborate with other >>> folks who are interested instead of duplicating the effort. >> >> There was a discussion about it at the European OpenZFS conference in Paris >> a couple of years ago. One of the drive vendors came along and gave a >> technical presentation on SMR, with a view to persuading ZFS developers to >> modify ZFS to make good use of SMR drives. >> >> The presentation was very good, but the general feedback at the time was >> that no one believed SMR would be around for very long, probably not long >> enough to get any support in ZFS stable, before hard drives had completely >> given way to SSDs. >> >> Things may have changed since then, although I hear less about SMR drives >> now than I did back then. How long do you think SMR will be around? What is >> your use case for the drives? > > Spectra’s take on it is that Seagate and HGST / WD will likely keep producing > high capacity SMR drives for a while. > > From a logical standpoint, SMR allows them to increase the drive capacity by > 20% or so (the numbers are fuzzy, that isn’t a precise amount), and they’ve > been selling SMR drives for a lower cost per GB than traditional fully random > access drives. > > SSD prices are coming down, and capacity is going up, but it still can’t > match hard drives in terms of price for the capacity. On the other end of > the spectrum, tape capacity and throughput is increasing pretty rapidly. > (LTO-8 is coming soon.) > > In between those is spinning disk. SMR disks have lower random write > performance, but also lower price, and are more competitive with tape. With > tape capacity increasing, the disk vendors will want to stay somewhat > competitive. SMR is a tool that they can use to compete on the lower price, > higher capacity end of the spectrum. > > Spectra’s focus is generally on archive and backup storage. That means high > capacity disk and tape libraries. Spectra generally only uses SSDs for > things like caching, databases, etc. They’re too expensive for archive > storage. > > As for Spectra’s use case for the drives: > > https://www.spectralogic.com/products/arcticblue/ > <https://www.spectralogic.com/products/arcticblue/> > > That’s a 96-drive, 4U JBOD enclosure. We can put up to 8 Arctic Blue > enclosures behind a Black Pearl (S3) box. Right now we’re using Drive > Managed SMR drives with ZFS. We only had to do some minor modifications to > ZFS to get reasonable performance with Drive Managed SMR drives. The cost > (as low as $.10/GB for the whole system) is very good. > > While one vendor has Drive Managed and Host Aware drives, the other one, as > far as I know, only has Host Managed drives. Host Managed is much harder to > deal with from a software standpoint, since it can’t handle any out of order > access within the bands. In order to have higher performance with Host Aware > drives, and be able to choose between vendors, we’ll have to implement Host > Managed support in ZFS, or write a layer below ZFS that will do > read/modify/write to eliminate any non-sequential writes to Host Managed > disks.
As part of the metadata classes discussion a while back, the feature would allow you to easily separate out the raw data from other metadata and place them on different vdevs. At that point, an SMR-optimized allocator makes sense for the data while metadata can reside on non-SMR drives. -- richard > > Ken > — > Ken Merry > [email protected] <mailto:[email protected]> > > > > openzfs-developer | Archives > <https://openzfs.topicbox.com/groups/developer/discussions/T1fec2160a70daccc-M63c1e798a0c2c95b02d82ed9> > | Powered by Topicbox <https://topicbox.com/> ------------------------------------------ openzfs-developer Archives: https://openzfs.topicbox.com/groups/developer/discussions/T1fec2160a70daccc-M07491d6c76fd073693b2e962 Powered by Topicbox: https://topicbox.com
