No, not a cricket question!! But one which is is nearly as important! We are now all coming to the happy agreement that wide-striping is generally a good thing! But just as there are some implementations of thin-provisioning which are a little chubbier than the others; there are implementations of wide-striping which are a little narrower than the others.
So at what point does striping become wide? Is 14 disks wide? Is it 16 disks? 26 disks? 50 disks? 100 disks? At what point, does the implementation become wide?
Any thoughts? Go on, I know there are!!
Wide Striping should have no limitations based on number of disks, in my opinion an array has “wide striping” capabilities if it can stripe data on ALL the disk you want, without any sort of limitations. On the other hand, if you consider an arbitrary number of disk as “enough” for wide striping, almost every vendor does it, for instance if i recall correctly HDS AMS line can stripe to 16+1 R5 on a single raid group and that’s a midrange storage…
just my 0.02 €
Wide means Wide: It’s not something about a big raid 5 group!
You can obtain a real wide striping only if you have a pool with unlimited number of disks, each i/o operation will be striped over all the disks!
Of course the pool need to be expandable with automatic rebalancing of the workload up to the limits of the array.
Whatever takes your fancy.
2 or 10,000 – your risk vs gain will make you decide whats best for your business.
I’ve heard this term a lot lately since I’ve been diving into storage technology but I’m not entirely sure what it is and Wikipedia doesn’t yet seem to have a page for it.
Question: is what EqualLogic does considered wide striping? by default the RAID spans all disks in an enclosure – I think this can be a bit limiting if I don’t want it but it seems to work well. If I add another array to an existing group, the group now spans the new disks and adds their controller, ethernet ports and disks to the original group… would this be considered wide striping? the disk part, obviously.
Just to pick up on one of the smaller points of your post – is wide-striping now definitely a good thing?
The trade-off with wide-striping is disk contention. Whereas before you may have had a couple of apps sharing a smaller number of disks, you now have all your apps striped across all your disks. The concensus appears to be that throwing as many resources (ie, spindles) at the problem (ie, total i/o / throughput) as possible and hoping that those resources are sufficient to cope, is a ‘good thing’.
Personally, I’m not a fan of the ‘throw resources at it till it’s no longer a problem’, but I can see why it’s appealing. Certainly from an administrative point of view, and in certain cases performance (see Claus’s blog entry: http://blogs.hds.com/claus/2009/06/anyone-interested-in-a-105000-rpm-drive.html )
The caveat on the performance front is that your array has to be heavily under-utilised to permit scope for those large performance increases.
We’ve employed wide-striping in the past (admittedly, a while back on older, slower kit) and have found that bursts of intense i/o from one app on a wide-striped group of disks crippled the performance of other apps sharing those disks.
Now, maybe we’ve entered a new paradigm and the problems we experienced are no longer a problem, but I’m not sure I believe that.
Anyway, we’ve got some new kit and scope to do wide-striping so we’re now going to do some testing over the next couple of weeks.
As for ‘how wide is wide’? Well, that’s easy. It’s just twice as many disks as half the original number :o)
In many ways I don’t disagree with you Jon but the problem is unless vendors come up with very superior tools to automatically balance arrays and spindles; arrays are simply too large to manually manage on a spindle by spindle basis to balance performance without incredibly large storage teams and very few of us have that luxury anymore.
Wide-striping is a quick and dirty fix which will work in 80-90% of the cases. It will get better with balancing which will monitor spindles and move data off to less loaded spindles or faster spindles. You still need to understand your application profiles BTW and if you have a known bursty application, you might want to handle it differently. Perhaps give it its own dedicated set of spindles; I mean you do performance test your applications before you put them live unless things have very much changed in the old place.
But I have seen some very impressive performance figures from engineering teams who have done their own testing with wide-striped arrays. Some of them in the same space as you.
I still suspect as drive sizes rise; utilisation rates will continue to fall; drives will be I/O saturated before capacity runs out.
Wide Striping “i think” is what EVA does as its basic striping method for raid groups. It works very well on EVA in smaller arrays.
Trouble is eggs in one basket I guess….also increased leveling times for new disks/raid groups.