[zfs-discuss] Issue with data

Joachim Grunwald jgrunwald at ecotronic.de
Thu Sep 26 05:28:52 EDT 2013


Am 25.09.2013 08:50, schrieb Richard Laager:
> On Wed, 2013-09-25 at 09:20 -0600, Marcus Sorensen wrote:
>> If he's right, then recordsize needs to be ashift in bytes * (n-1) for
>> raidz.
> For space efficiency, you want:
> recordsize >= 2^ashift * (# of data disks)
What do you mean by recordsize (=filesize)?
What influence do I have on the recordsize if I am writing directly on 
the pool?
>
> Avoiding this surprise is part of the thinking behind powers-of-two
> numbers of data disks in raidz* as well.
>
> Note that things still aren't perfect in the real-world because you're
> almost always going to use compression with ZFS (typically lz4 these
> days). However, compression can still only ever save space, so it
> doesn't lead to the surprises over raidz space-inefficiency that I see
> from new users in IRC from time to time.
>
>> I would have expected zfs to spread the write out over a full stripe
>> even if you did something like this, maybe writing 2k to each disk
>> (you'd still lose 50% space if your a shift was 12 in this case).
> Why do you want ZFS to use 5*4K instead of 3*4K? That's simply wasting
> disk space; and worse, when reading data, you now need to read from 5
> disks instead of only 3, which reduces your overall IOPS even further.
>


To unsubscribe from this group and stop receiving emails from it, send an email to zfs-discuss+unsubscribe at zfsonlinux.org.



More information about the zfs-discuss mailing list