Specific NAS requirements (24x7 write 100-200MB/s and sometimes little read 1-50MB/s)

[danb35]

Would you buy 128TB when you need 100TB ?

When he says "any reserve is appreciated", yes. That says (at least to me) that the 100 TB he mentions isn't a hard upper bound, and more space would be beneficial.

IcePlanet was planning to have a Quad 1Gb NIC on the NAS and connect each blackbox to a dedicated NIC.

Where did you see him say that? Because I didn't see it. I saw the quad NIC, but no statement that each blackbox would be directly connected to a dedicated port on it.

 

[conceptual_continuity]

Hi,



The point is about the requirements and you are mixing iops and throughput.

200Mb/s, Writing, Big files, No synchronised writes so anyway the iops will be provided by the cache.
Where is the iops requirement ? This is about throughput here.

There is not point in justificating a high number of vdev with performance. This results in a waste of space.



Why do you think I am not aware of this rule ?

160TB*0.8=128TB, not 100TB.
Would you buy 128TB when you need 100TB ?
Added to (1), this is quite a lot of wasted disks.


I don't think 1G network link will transfer 1Gbit/sec. Why may I think that ?
What is your justification for "With the OP requested guaranteed speed, you have to go at least with 10G" in this particular case ?

The 1GB is on the blackbox side.
IcePlanet was planning to have a Quad 1Gb NIC on the NAS and connect each blackbox to a dedicated NIC.
So no 10Gb NIC here and actually this might be the only specification that make sense in his original post.

So yes ok, all this is nice, but where is the point for this NAS ?

Thanks for your reply, Thomas.

IOPS is crucial, if not in the beginning (unless you never delete stuff of course!). At some point fragmentation will be a limiting factor, and you're IOPS limited. That's why we read posts like "suddenly my pool turned slow...".
The write caching in RAM (in case of non-sync writes) won't help if the disk heads cannot keep up... we have *continuous* write load.

As far as space goes: 160TB hold about 144 TiB (Terabyte) which corresponds to filling the pool with ~70% of data. The 80% rule of thumb already means considerably reduced write performance (do not have the link right now), so if sustained high speed write performance over years is *important*, I would go with that slight overprovisioning.
Disk space is relatively cheap, so what.
Then again, going with many vdevs means you can start out with just a few while fragmentation is still low, and keep them adding as your pool fills up and fragmentation grows.

Regarding the NICs: why the cludge of 4 1G links when a 10G connection is not much more expensive these days and you can easily satisfy additional requirements (e.g. read) traffic. The OP wanted to be able to take drives out to connect them to a PC - as this won't work he'll need the fast network connection for an alternative strategy anyway.

Apart from that, I would lose some thoughts on making sure Samba (single threaded) does not become a bottleneck CPU wise.

Good luck to the OP

 

[Thomas102]

His danb

When he says "any reserve is appreciated", yes. That says (at least to me) that the 100 TB he mentions isn't a hard upper bound, and more space would be beneficial.

Yes, so it is not related to the 80% limit conceptual_continuity spoke about.

Where did you see him say that? Because I didn't see it. I saw the quad NIC, but no statement that each blackbox would be directly connected to a dedicated port on it.

For me, this is explicit enough. 4 ports NIC, 4 items to connect, :

one 4xGBIT card to connect the blackboxes and GBIT switch for read operations
[/QOTE]

 

[Chris Moore]

It's not a good solution to direct connect network nodes, even if the OP was thinking about that, it is another place where the correct answer is different than the OP question.

Sent from my SAMSUNG-SGH-I537 using Tapatalk