[Gluster-users] Design/HW for cost-efficient NL archive >= 0.5PB?

Thu Jan 2 19:19:02 UTC 2014

Sent from my iPad

On 02.01.2014, at 18:06, Justin Dossey <jbd at podomatic.com> wrote:

> 1) It depends on the number of drives per chassis, your tolerance for risk, and the speed of rebuilds.  I'd recommend doing a couple of test rebuilds with different array sizes to see how fast your controller and drives can complete them, and then comparing the rebuild completion times to your SLA-- if a rebuild takes two days to complete, is that good enough for you (especially given the chances of another failure occuring during the rebuild)?  All other things being equal, the smaller the array, the faster the rebuild, but the more "wasted" space in the array.  Also note that many controllers have tunable rebuild algorithms, so you can divert more resources to completing rebuilds faster at the cost of performance.  One data point from me: my last 16-2T-SATA RAID-6 rebuild took about 58 hours to complete.
> 
> 2) My understanding is that the way file reads work on GlusterFS, read requests are sent to all nodes and the data is used from the first node to respond to the request.  So if one node is busier than others, it is likely to respond more slowly and thus receive a lower portion of the read activity, as long as the files being read are larger than a single response.    
> 
> 
> On Wed, Jan 1, 2014 at 12:21 PM, Fredrik Häll <hall.fredrik at gmail.com> wrote:
>> Thanks for all the input!
>> 
>> It sure sounds like RAID-6 for disk failures and Gluster for the spanning and high level redundancy parts is a good candidate. 
>> 
>> Some final questions: 
>> 
>> 1) How big can one comfortably go in terms of RAID-6 array size? Given 4TB SATA/SAS drives. On the one hand much points to keeping as few RAIDs as possible, and disk usage is of course maximized. But there are complications in terms of rebuild times and risk of losing the 2 drives. Hot spares may also be an option. Your reflections?
>> 
>> 2) Is there any intelligence or automation in Gluster that makes smart use of dual (or multiple) replicas? Say that I have 2 replicas, and one of them is spending some effort on a RAID rebuild, is there functionality for manually or automatically preferring the other (healhy) replica?
>> 
>> Best regards, 
>> 
>> Fredrik
>> 
>> 
>> On Tue, Dec 31, 2013 at 10:27 PM, Justin Dossey <jbd at podomatic.com> wrote:
>>> Yes, RAID-6 is better than RAID-5 in most cases.  I agonized over the decision to deploy 5 for my Gluster cluster, and the reason I went with 5 is that the number of drives in the brick was (IMO) acceptably low.  I use 6 for my 16-drive arrays, which means I have to lose 3 disks out of the 16 to lose my data.  With 2x8-drive arrays in 5, I also have to lose 3 disks to lose data, but if I do lose data, I only lose 50% of the data on the server, and all these bricks are distribute-replicate anyway, so I wouldn't actually lose any data at all.  That consideration, paired with the fact that I keep spares on hand and replace failed drives within a day or two, means that I'm okay with running 2x RAID-5 instead of 1x RAID-6.  (2x RAID-6 would put me below my storage target, forcing additional hardware purchases.)
>>> 
>>> I suppose the short answer is "evaluate your storage needs carefully."
>>> 
>>> 
>>> On Tue, Dec 31, 2013 at 11:19 AM, James <purpleidea at gmail.com> wrote:
>>>> On Tue, Dec 31, 2013 at 11:33 AM, Justin Dossey <jbd at podomatic.com> wrote:
>>>> >
>>>> > Yes, I'd recommend sticking with RAID in addition to GlusterFS.  The cluster I'm mid-build on (it's a live migration) is 18x RAID-5 bricks on 9 servers.  Each RAID-5 brick is 8 2T drives, so about 13T usable.  It's better to deal with a RAID when a disk fails than to have to pull and replace the brick, and I believe Red Hat's official recommendation is still to minimize the number of bricks per server (which makes me a rebel for having two, I suppose).  9 (slow-ish, SATA RAID) servers easily saturate 1Gbit on a busy day.
>>>> 
>>>> 
>>>> I think RedHat also recommends RAID6 instead of RAID5. In any case, I
>>>> sure do, at least.
>>>> 
>>>> James
>>>> 
>>>> 
>>>> 
>>>> On Mon, Dec 30, 2013 at 5:54 AM, bernhard glomm
>>>> <bernhard.glomm at ecologic.eu> wrote:
>>>> >
>>>> > some years ago I had a similar tasks.
>>>> > I did:
>>>> > - We had disk arrays with 24 slots, with optional 4 JBODS (each 24 slots) stacked on top, dual LWL controller 4GB (costs ;-)
>>>> > - creating raids (6) with not more than 7 disks each
>>>> > - as far as I remember I had one hot spare per each 4 raids
>>>> > - connecting as many of this raid bricks together with striped glusterfs as needed
>>>> > - as for replication, I was planing for an offside duplicate of this architecture and
>>>> > because losing data was REALLY not an option, writing it all off at a second offside location onto LTFS tapes.
>>>> > As the original version for the LTFS library edition was far to expensive for us
>>>> > I found an alternative solution that does the same thing
>>>> > but fort a much reasonable prize. LTFS is still a big thing in digital Archiving.
>>>> > Give me a note if you like more details on that.
>>>> >
>>>> > - This way I could fsck all (not to big) raids in parallel (sped things up)
>>>> > - proper robustness against disk failure
>>>> > - space that could grow infinite in size (add more and bigger disks) and keep up with access speed (ad more server) at a pretty foreseeable prize
>>>> > - LTFS in the vault provided just the finishing having data accessible even if two out three sides are down,
>>>> > reasonable prize, (for instance no heat problem at the tape location)
>>>> > Nowadays I would go for the same approach except zfs raidz3 bricks (at least do a thorough test on it)
>>>> > instead of (small) hardware raid bricks.
>>>> > As for simplicity and robustness I wouldn't like to end up with several hundred glusterfs bricks, each on one individual disk,
>>>> > but rather leaving disk failure prevention either to hardware raid or zfs and using gluster to connect this bricks into the
>>>> > fs size I need(  - and for mirroring the whole thing to a second side if needed)
>>>> > hth
>>>> > Bernhard
>>>> >
>>>> >
>>>> >
>>>> > Bernhard Glomm
>>>> > IT Administration
>>>> >
>>>> > Phone: +49 (30) 86880 134
>>>> > Fax: +49 (30) 86880 100
>>>> > Skype: bernhard.glomm.ecologic
>>>> > Ecologic Institut gemeinnützige GmbH | Pfalzburger Str. 43/44 | 10717 Berlin | Germany
>>>> > GF: R. Andreas Kraemer | AG: Charlottenburg HRB 57947 | USt/VAT-IdNr.: DE811963464
>>>> > Ecologic™ is a Trade Mark (TM) of Ecologic Institut gemeinnützige GmbH
>>>> > ________________________________
>>>> >
>>>> > On Dec 25, 2013, at 8:47 PM, Fredrik Häll <hall.fredrik at gmail.com> wrote:
>>>> >
>>>> > I am new to Gluster, but so far it seems very attractive for my needs. I am trying to assess its suitability for a cost-efficient storage problem I am tackling. Hopefully someone can help me find how to best solve my problem.
>>>> >
>>>> > Capacity:
>>>> > Start with around 0.5PB usable
>>>> >
>>>> > Redundancy:
>>>> > 2 replicas with non-RAID is not sufficient. Either 3 replicas with non-raid or some combination of 2 replicas and RAID?
>>>> >
>>>> > File types:
>>>> > Large files, around 400-1500MB each.
>>>> >
>>>> > Usage pattern:
>>>> > Archive (not sure if this matches nearline or not..) with files being added at around 200-300GB/day (3-400 files/day). Very few reads, order of 10 file accesses per day. Concurrent reads highly unlikely.
>>>> >
>>>> > The main two factors for me are cost and redundancy. Losing data is not an option, being an archive solution. Cost/usable TB is the other key factor, as we see growth estimates of 100-500TB/year.
>>>> >
>>>> > Looking just at $/TB, a RAID-based approach to me sounds more efficient. But RAID rebuild times with large arrays of large capacity drives sound really scary. Not sure if something smart can be done since we will still have a replica left during the rebuild?
>>>> >
>>>> > So, any suggestions on what would be possible and cost-efficient solutions?
>>>> >
>>>> > - Any experience on dense servers, what is advisable? 24/36/50/60 slots?
>>>> > - SAS expanders/storage pods?
>>>> > - RAID vs non-RAID?
>>>> > - Number of replicas etc?
>>>> >
>>>> > Best,
>>>> >
>>>> > Fredrik
>>>> > _______________________________________________
>>>> > Gluster-users mailing list
>>>> > Gluster-users at gluster.org
>>>> > http://supercolony.gluster.org/mailman/listinfo/gluster-users
>>>> >
>>>> >
>>>> >
>>>> > _______________________________________________
>>>> > Gluster-users mailing list
>>>> > Gluster-users at gluster.org
>>>> > http://supercolony.gluster.org/mailman/listinfo/gluster-users
>>>> 
>>>> 
>>>> 
>>>> 
>>>> > --
>>>> > Justin Dossey
>>>> > CTO, PodOmatic
>>>> >
>>>> >
>>>> > _______________________________________________
>>>> > Gluster-users mailing list
>>>> > Gluster-users at gluster.org
>>>> > http://supercolony.gluster.org/mailman/listinfo/gluster-users
>>> 
>>> 
>>> 
>>> -- 
>>> Justin Dossey
>>> CTO, PodOmatic
>> 
>> 
>> _______________________________________________
>> Gluster-users mailing list
>> Gluster-users at gluster.org
>> http://supercolony.gluster.org/mailman/listinfo/gluster-users
> 
> 
> 
> -- 
> Justin Dossey
> CTO, PodOmatic
> 
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://supercolony.gluster.org/mailman/listinfo/gluster-users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20140102/6ae495a7/attachment.html>