Matthias Andree: > On Fri, 18 Feb 2011, Wietse Venema wrote: > > > Please file a ZFS bug reportug. As per POSIX, when the O_CREAT is > > specified to open(), > > > > The third argument does not affect whether the file is open > > for reading, writing or for both. > > > > In other words, read/write access is controlled with the O_RDWR flags, > > not the read/write permissions argument. > > > > When the above error happens, Postfix discards the file. Consequently, > > Postfix performance will be reduced, because it creates one extra file > > per MAIL transaction, instead of one extra file over the process > > lifetime. > > That being said, I found ZFS performance abysmal, that is with a > single-disk pool (one partition on an enterprise-class 7200/min HDD, > amd64, 4 GB DDR3 ECC RAM, Phenom II X4 905e i. e. 4 x 2.5 GHz) on > FreeBSD 8.X, and ZFS prefetching enabled in the loader tunable. > > I'd been running /usr with compression enabled, and stat() effort, > writes, and thereabouts were unbearably slow, even with a 85% filled > partition (the pool was 100% full one time before). I've switched to > UFS2 on a 5400/min eco-consumer-class drive for the nonce and it just > flies. I haven't analyzed this, but it seems there still are a few > rough edges in ZFS-on-FreeBSD that need to be filed down a bit :)
Let me guess: the eco-consumer-class drive was running with write caching enabled. I remember achieving better latencies on a $500 computer than with a (small) server whose disk alone cost hundreds. > Note that ZFS on FreeBSD also still has a bug that it can't delete files > if the partition is full. You need to manually truncate files (to create > room for the relevant deletion log entries) before you can remove files. > (I've file a problem report for that but don't know the PR # yet.) Will ZFS allow non-root users to fill up the whole disk? Wietse