One of the lessons learned by oracle was that you don’t want to just run around trimming things - they did end up with adding minimal suggested amount to trim, otherwise the storage did suffer… I may be mistaken, but something like 1MB or so. Another question is, the trim is not just about ssd, but also about thin provisioned storage from arrays (where the excessive amount of small releases can cause even more damage;)?
Perhaps trim/autotrim could use pool property for minimal trim size (sorry, havent had yet time to read into the review…). rgds, toomas > On 7. aug 2016, at 23:31, Saso Kiselkov <[email protected]> wrote: > > On 8/7/16 8:07 PM, Garrett D'Amore wrote: >> historically some ssds have performed worse with trim as they do >> some expensive operations. i have not looked at the webrev yet >> but are there tunables or an administrative interface to this? > > By default, everything is off, so there are no performance > implementations "out of the box". The administrative interfaces are, either: > > * enabling automatic trimming after deleting data: > # zpool set autotrim=on <poolname> > > * manually initiating a single full-pool trim (running through > all the free space on a pool and trimming it): > # zpool trim <poolname> > > Automatic trim is as the name implies: fully automatic. You just delete > data and ZFS will at some later point, if the blocks have not been > reused, simply trim those blocks out. > > Manual trim behaves much like a scrub. It has an adjustable rate (in > bytes/sec) and can be monitored through "zpool status" or stopped at any > time via "zpool trim -s <poolname>". > >> how diverse in terms of ssds has this been tested with? > > We've tested primarily with Sandisk SAS SSDs, mainly the Optimus line > and on some pretty high-end kit (Sandisk's INFINIFLASH). > >> are there performance implications observed? > > Yes, this is detailed in the manpage. In general, Optimus 1 takes poorly > to automatic trim and can suffer a write throughput limitation when lots > of deletes are going on. On Optimus 2 and the IF100 the performance > impact was much smaller (measurable, but really not much). > > In general, the manpage suggests a "test and see what works" approach. > Try autotrim=on and see if it affects performance on your system. If it > doesn't, leave autotrim on. If it does, disable autotrim and use manual > trim during a maintenance window to still get most of the used block > recovery benefits. > > Cheers, > -- > Saso > ------------------------------------------- openzfs-developer Archives: https://www.listbox.com/member/archive/274414/=now RSS Feed: https://www.listbox.com/member/archive/rss/274414/28015062-cce53afa Modify Your Subscription: https://www.listbox.com/member/?member_id=28015062&id_secret=28015062-f966d51c Powered by Listbox: http://www.listbox.com
