Can't Delete Files From Full Drive

[joeschmuck]

Would quotas work? Or does a dataset with a quota have the same issue if you fill up to the quota limit? I haven't hit a quota limit yet, but I broke up my datsets with quotas that add up to less than 80% of the pool to keep myself from accidentally filling it.

Yes, so long as you do not go below the actual 10% fill for your pool.

I thought once you exceeded 20% fragmentation occur in the longer run which can permanently slow the pool..

ZFS by nature is fragmented, all the pieces of a file are tossed all over the pool including parity. Once you hit 10% (or maybe it's just below 10%) the ZFS starts to use a different method to place data on the drive in an effort to ensure it all fits properly and this will slow down your system, sometimes considerably.

 

[ser_rhaegar]

Thanks Joe

 

[Yatti420]

Yes, so long as you do not go below the actual 10% fill for your pool.


ZFS by nature is fragmented, all the pieces of a file are tossed all over the pool including parity. Once you hit 10% (or maybe it's just below 10%) the ZFS starts to use a different method to place data on the drive in an effort to ensure it all fits properly and this will slow down your system, sometimes considerably.

The ratio is 20% isn't it? After 20% if goes to best available versus normal or something?

 

[cyberjock]

No, it's at 95%. Hard coded by FreeBSD from what I've read. It used to be 80%, then 90% with Solaris.

 

[joeschmuck]

So FreeBSD is 95% full, damn that is the first time I've heard that one because I know we were all talking 10% remaining probably a year ago. Now I'm going to look that up somehow just so it sticks into my head and find out if that was new for FreeBSD 9 or has been that way since FreeBSD 8.

EDIT: So I found the reference to 4% remaining being the trigger point from first-fit to best-fit. Guess there was a lot of confusion between people talking Oracle, Solaris, and FreeBSD.

 

[cyberjock]

Oh yeah.. lots of stuff we talk about here is either hard to validate or wrong because there's differences between Oracle, Solaris, FreeBSD, ZFSonLinux, etc. So the source has to be considered. Unfortunately, if that source got its source from a bad place, then it's still wrong. Sucky? Absolutely.

 

[warri]

Yep, 4% seems to be the value used by FreeBSD. I always thought it was closer to 15.. - note to self: double check random stuff you read on the internet ;)

For a reference see here in the source code.

And indeed, FreeNAS didn't change this default:

Code:

[root@freenas] ~# sysctl vfs.zfs.metaslab.df_free_pct
vfs.zfs.metaslab.df_free_pct: 4