[CentOS] ZFS

Mon Sep 15 21:40:50 UTC 2014
Warren Young <warren at etr-usa.com>

On 9/15/2014 14:48, Andrew Holway wrote:
>>
>>   Any comparison between ZFS and non-ZFS probably overlooks things like
>> fully-checksummed data (not just metadata) and redundant copies.  ZFS will
>> always be slower than filesystems without these features.  TANSTAAFL.
>
>
> Not really true. It hugely depends on your workload. For example, if you
> have a 20TB filesystem with 128GB of "ARC" (adaptive replacement cache)
> then ZFS will be many many times faster then ext4 assuming that the "hot
> data" is under 128GB as all reads will come from memory or a dedicated
> cache SSD. If however you are streaming the whole 20TB from the filesystem
> then the cache makes no difference and you just see the performance of the
> disks. The checksumming for example does not typically add a performance
> penalty as this calculation is done in parallel to normal disk operations.

The SSD and second CPU core are not free.

ZFS does more than pretty much every other mainstream filesystem, and 
hence needs more resources when you use its additional features.

My only point is that there are people who will compare a ZFS pool with 
all the features turned on to a bare-bones ext4 setup, then complain 
that ZFS is "slow."

> The atomistic write mechanism of ZFS can be hundreds of times faster than
> EXT in cases where your IO is random SYNC assuming you have a dedicated ZIL
> (ZFS intent log) which keeps the journal. Read up on ZFS transaction groups
> for more details.

The "ZFS is slow" crowd generally doesn't care about reliable fsyncs. 
Hence why most PC OSes lie through their teeth when you tell them to 
sync a block to disk:

     http://queue.acm.org/detail.cfm?id=2367378