[zfs-discuss] Re: Fragmentation vs traditional linux FS and ZFS as a datastore for SQL db's

Gordan Bobic gordan.bobic at gmail.com
Sat Jan 18 06:59:44 EST 2014

On 01/18/2014 10:48 AM, Andrew Holway wrote:
>> If one needs performance, one will go with "persistent memory" kind of
>> storage, like Fusion-IO's IOdrive, with latencies in the order of tens of
>> microseconds.
> Which is PCIe and not SAS.
> Running your database on a PCIe card or even a SATA SSD is great and
> super quick but what happens when the machine breaks.

You fail over the the replicated mirror.

> For "traditional
> style" mission critical stuff people will tend to use SAS drives as
> their dual port magic enables you to fail the storage over to another
> host.

What happens when your storage fails? Say you have a failure on the 
power supply/bus powering those dual-ported disks via their single power 

If you are serious about redundancy, you have two of everything, which 
means a complete second server with a complete second set of disks and 
data replication between them.

> SAS SSD are however still a little thin on the ground which
> makes it tricky to use them as persistent storage for these kinds of
> applications.
> For instance the whole Intel SSD range is still SATA.
> Also......duh......write wearing.

Before you cry "write endurance", do sit down and do some maths 
comparing the actual observed write-throughput on a system with the 
write endurance of the underlying devices. For 99% of deployments write 
endurance will never be an issue. I have only ever seen one production 
system where I would be even remotely concerned about putting the data 
on SSDs, and even then it entirely depends on what SSD you put them on.

The production system in question was handling a throughput of about a 
PB/week with a storage capacity of about 8TB. Intel S3700 is rated for 
10 complete capacity overwrites per day for 5 years, so the 800GB model 
is good for 14.6PB over it's lifetime. We'd need 10 of those for the 8TB 
capacity, so we are lookig at 146PB of write endurance across the pool. 
At 1PB/week, that is about 3 years. With most business uses you'd be 
upgrading your servers, or at least your disks for extra capacity after 
3 years anyway.

And that case of 1PB/week is a pretty pathological worst-case for 
meaningful production database use.

So unless my math above is out by an order of magnitude somewhere, flash 
write endurance is a non-issue even for the most pathological of loads.


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