ZFS regimen: scrub, scrub, scrub and scrub again.
Wojciech Puchar
wojtek at wojtek.tensor.gdynia.pl
Wed Jan 23 22:52:38 UTC 2013
>> unless your work is serving movies it doesn't matter.
>
> That's why I find it really interesting the Netflix Open Connect
> appliance didn't use ZFS - it would have seemed perfect for that
"Seems perfect" only by ZFS marketers and their victims.
but is at most usable, but and dangerous.
> application.
because doing it with UFS is ACTUALLY perfect.
large parallel transfers are great with UFS, >95% of platter speed is
normal and near zero CPU load, metadata amount are minimal and doesn't
matter for performance and fsck time (but +J would make it even smaller).
Getting ca 90% of platter speed under multitasking load is possible with
proper setup.
> http://lists.freebsd.org/pipermail/freebsd-stable/2012-June/068129.html
>
> Instead there are plain UFS+J filesystems on some 36 disks and no RAID -
> it tries to handle almost everything at the application layer instead.
this is exactly the kind of setup i would do in their case.
They can restore all data as master movie storage is not here.
but they have to restore 2 drives in case of 2 drives failing in the same
time. not 36 :)
"application layer" is quite trivial - just store where each movie is.
such setup could easily handle 2 10Gb/s cards. Or more if load is spread
over drives.
More information about the freebsd-fs
mailing list