Like a certain Alex McDonald, I don’t get spin-down! Okay, I get the idea; when a disk isn’t being accessed, after a certain period of time you can spin it down and hence you can save power, wave your little eco-worrier flag and feel good about yourself. But does it work?
Okay, first things first; I do not believe that spin-down fixes one of the fundamental challenges that a lot of data centre managers face today; they simply cannot get enough power into their data centres. Their data centres may be in the wrong place but that’s not an easily solvable issue without building new ones (and there isn’t alot of appetite for that at this moment). So they need to engage their local power-companies to upgrade the power coming into the buildings; not a cheap options.
But hang-on a minute, surely spun-down disks require less power; actually no, not really, they need less power on average but you still need to be able to cope with the peak-load which is everything spun-up. It might not be a likely scenario but if you can’t supply enough to power to the spin them all up; you might find some interesting problems. I’ve been through the fun and games of having to shut things down extremely quickly due to over-loaded and over-heated PDUs; ones with smoke pouring out of them and electricians with fire extinguishers trying to buy us time to do clean shutdowns. So I still need to upgrade my power.
And then, wide-striping is going to completely stuff spin-down. So we’ll end up with a non-wide-striped pool; I’d even suggest that you’d probably want to avoid parity RAID and stick with RAID-1.
Making spin-down work in a block-environment is going to be a challenge; volume managers are probably going to spin all the disks up in a volume. I guess a 1-to-1 relationship between disk and volume would work but I reckon most of the disks would still end up spinning. Or I could define spin-down pools I guess and use some kind of stub to point to the migrated files.
Spin-down in NAS environments may be more practical and may be easier to implement.
And yes, it works pretty well with desk-tops (mostly avoids crashing my computer).
But I’m really struggling at the moment to see how it’s going to work at the moment; I suspect I need to do some more research but some pointers would be nice! And what’s wrong with tape? It’s the ultimate spun-down magnetic rust!
Have you looked at Copan MAID?
Only ever spins up 25% of the total number of disks.
Copan…indeed!!
Power-managed RAID!!
And yes, you are correct that they guarantee that only 25% of the disks spin up. How do they do this? If a file sits on a powered-down disk, it powers down an operational one and powers up the powered-down disk. This means potentially significantly increased latency and means work on the applications.
I wonder, do I not get better savings by going to dedupe on primary storage; just a thought?
I’m just yet to be convinced I guess.
Linkee at the top no workee, so here it is again; http://blogs.netapp.com/shadeofblue/2008/11/lets-guess-does.html
I’m going to write a set of posts on storage costs; the lack of a cost model to set against technlogies like spin down, SSD, and verious other “save you money” claims needs tested.
On hols this week, so not until next.
Link fixed, thanks!!
Enjoy your holiday!!
Martin
One thing to consider. What are the chances that your data is contained on just the drives you intend to spin down? If the data is on RAID, it could be 2-28 disks wide. Then there’s thin provisioning, wide striping, all of which potentially contend with spin down.
Completely agree Chris. Spin-down could be another tier which will need some special rules/technology. Parity RAID, probably not…dispersed object-based storage, could be. But it’s not an instant panacea and it needs some careful thought, more so than I think it is sold with.
Perhaps it’s these implementation details which has left Copan with relatively few customers for a company of it’s age.
I remember back in 2004 getting righteously excited about an academic paper I’d found off USENIX describing MAID.
At the time, I had been asked to figure out a way to store the whole of a large media company’s video archive online. That should have been a simple enough problem (the answer was DON’T – i.e. use tape) but that was the problem I had to solve. Then, and now, a well implemented spin-down solution can be useful in that kind of environment. You have very large quantities of relatively infrequently accessed stuff (the highest quality version of content), much of which doesn’t need instant access but which should be accessible ‘quickly’ – i.e. quicker than tape. You also have comparatively few objects per TB, which again reduces the likelyhood that any given RAID group is going to end up spun-up. Finally, in the media world we just love our proxies so in the end, we very infrequently hit the big stuff.
Outside of use-cases like that where management is relatively simple because you have a single app, the storage admin’s biggest challenge is the same they face for tiering etc. That’s to understand the applications, which are running on their tin and have service levels understood by the people who own those applications.
Had an interesting conversation with Brad Rigden at the BBC about tape and the confusion that people get into when talking about media going tapeless. And that we don’t normally mean tapeless, we mean digital tape and actually we can do online tape pretty near to disk-type latencies for broadcast/media application. Okay, you keep your browse-proxy on spinning rust but tape is not as bad as most people think.
Heh. IT tape -v- video tape is an eternal problem. In the work I’ve done since Auntie, I’ve found that saying ‘file-based’ rather than tapeless gets the message across. Personally I hate ‘file-based’ too as it implies that files are the be-all and end-all which is definitely not something I agree with (a subject for a blog post one day I think).
What do you think of Aunties media desktop initiatives (I forget the official name)? I’ve still got a bunch of friends at SBS busy delivering it.