Now I like Chuck’s blog; its often informative but also often controversial; upsetting the other vendors left, right and centre and I suspect he might have done it again with this post here; it all starts nice and informative, discussing use cases for EFDs and especially Oracle, how it may change things and how to get the most out of EFDs, you might want to design for I/O hotspots to enable you to drive your EFDs as much as possible. No more spreading I/O intensive workloads across disks which is wasteful, time consuming and complex.
Nothing especially new here, I’ve also been talking about sub-lun paradigms with a few vendors; the fact that even in your hottest files, you may well find that only 10% of that file is actually hot and really it would be better if you could get that 10% onto EFD and not the whole file. However, this is going to need really good optimisation code, it’s going to get almost too hard for an individual to do. So nothing especially new and no startling revelations yet; especially none on how EMC are really going implement this.
And then Chuck can’t resist having a pop at the ‘spindle randomizers’ i.e NetApp, XIV, EVA et all; how are you going to do this? I’m not sure how they are all going to do it either but as I say, I currently can’t see how EMC are going to do it either at the moment.
I have some ideas how it could be done which I might go into in another post when I’ve got some more time to think it through.
About the sub-lun paradigm you should check out compellent (www.compellent.com) and their data progression / fast track combo, they do a lot of interesting stuff on their storage.
I’ve looked at Compellent in the past and yes, it certainly looks interesting. There’s a whole other post brewing on the usefullness of LUNs bubbling up at the moment; put it like this, I think LUNs are loosing their usefullness as an abstraction.
Hi, it’s the chief instigator at it again.
One way of thinking about this problem comes from the past — if you’re familiar with how a Symmetrix uses its nonvolatile cache, it’s roughly a similar problem, isn’t it?
There, the abstraction is “track” rather than “LUN”, e.g. a finer grained unit of optimization.
And, no, I’m not going to say what we’re gonna do about it, I’m just pointing out what’s been done in the past 🙂
Ahhh, the days of calculating storage in tracks and cylinders etc; that takes me back!!
Solve it all by virtualizing and – taking Martin G back again – use HSM to make sure the right stuff is on the right stuff…
Putting it all behind a single monolith is all well and good, but getting it out of there…