I'm currently working on a hybrid approach which dynamically switches between first-fit to best-fit and vice-versa. With this new allocator you could choose to always do best-fit if your workload requires something like that. You should see this in Opensolaris in the next month.
Thanks, George Jeff Bonwick wrote: > Yes. > > The ZFS block allocator has three independent layers: device selection, > metaslab selection (within the chosen device), and block selection > (within the chosen metaslab). > > Device selection is currently round-robin to provide dynamic striping, > with weighting for load balance. Many other policies are possible. > > Metaslab selection is based on the metaslab's weight, as computed by > metaslab_weight(). This considers the recording zone, locality, and > other factors. Again, the weight function could literally be anything. > > Finally, block selection is determined by the space map ops vector. > We currently have just one (a modified first-fit allocator), but we > have a couple of new ones in the works. It should be straightforward > to test-drive your own using metaslab_ff_alloc() as an example. > (The current first-fit allocator is just 70 lines of code.) > > Jeff > > On Mon, Mar 23, 2009 at 01:25:12PM -0700, David Dew wrote: > >> Hello, looking to see if it is possible to change the allocation strategy >> to a "best-fit" or "fill, spill" model. In looking into the current code >> and blog entries, there seems to have been some of this flexibility in mind. >> My current thoughts are this should be attempted in metaslab.c in the >> metaslab_alloc_dva function, but curious if anyone else has attempted this. >> >> Thanks >> -- >> This message posted from opensolaris.org >> _______________________________________________ >> zfs-code mailing list >> zfs-code at opensolaris.org >> http://mail.opensolaris.org/mailman/listinfo/zfs-code >> > _______________________________________________ > zfs-code mailing list > zfs-code at opensolaris.org > http://mail.opensolaris.org/mailman/listinfo/zfs-code >
