Slashdot Mirror
IBM Developing Lego-like Storage Brick
AaronW writes "According to this story at EE Times IBM is developing a 32TB storage system built around blocks that can be stacked like Lego bricks. Apparently they will be connected in a 3x3x3 mesh using capacitive coupling and will be water cooled."
This is just a marketing ploy so they can sell storage Clusters shaped like Castles, Pirate Ships and the Millenium Falcon!
No I am serious, can you just imagine it?
>>>>>> Chewie, take the professor in the back and plug him into the hyperdrive.
"IBM's Ice Cube project aims to define a way for end users to easily maintain increasing amounts of data, while also plowing ground for a similar approach to computing systems."
Ice Cube? Lemme guess: They sell a bandwidth package for Internet hosting called "Ice T"
Their bandwidth monitoring and packet sniffer is called "Snoop Dog."
Oh wait...IBM's PS/2 had the MCA bus. Maybe that was a Beastie Boys reference. Maybe IBM has been into Rap and the like for a long time...
Am I wrong in thinking that this design may lead to a new approach to servers farm, where each cube offers some kind of power (processing, storage, networking, moka brewing), and the whole assembly keeps itself in shape?
/., the assertion "Imagine a cluster of these!" takes its full meaning: storage might be the first step, and only the bandwidth of the couplers is a limit to the usability of CPU cubes or networking cubes.
For the first time in the history of
More, the software part will certainly bring some huge advances in clustering, as the challenge of virtualising all those cubes may help in building self-repairing (or should I say self-dumping?) clusters...
Oh, and by the way, here is the first step to assimilation.
--
Arkan
I'm not sure I follow. They say they want it to be easily stackable, and fault tolerant (they specifically mention leaving blocks in place if they fail), but how do you combine that with a water cooling system?
With a water cooling system, you need to make sure that the joints between cubes are water-tight, and maintain them over time, thus defeating their "no maitance" theory.
Or am I missing something? Perhaps they could use "disk blocks" and "cooling blocks" and just swap out the "cooling blocks" if there is a problem? Still takes more work than air cooling, but less than inegrating it into every block would.
What about just leaving air holes, and using it in a chilled room? Most server rooms are chilled anyway.
Just some ignorant thoughts.
Colin Davis
Colin Davis
If I were IBM I would avoid Lego comparisions
and 2nd, I would change the name.
Chicago2600.net more than a lifestyle, its a survival trait.
2. Tinkertoy: storage structures too delicate, engineers kept losing fins for making "windmill" structure.
3: Play-Doh: kept getting stuck in carpet.
4. Erector Set: engineers spent too much time making jokes about name.
See Robert Morris's presentation (6+MB PDF) from the USENIX File and Storage Technologies conference. The videos of the invited talks are also worth watching (if you can afford the b/width to get them).
...do you end up having to pull em apart with your teeth? I'd rather not, I'm sure they get really hot.
"Your superior intellect is no match for our puny weapons!"
Looks like an excellent step towards a truly borg like information technology system.
IBM: resistence is futile!
Websurfing done right! StumbleUpon
augment your senses: http://sensebridge.net/
I'm glad they finally announced this project I've been dying to talk about it. I talked to a researcher on this project while I was at IBM's Almaden Research Center.
I was blown away when they described it to me. I have to say that IBM is by far the greatest computer techonology research company. They take the top minds give them boat loads of money, ten years later they blow your mind with the completely innovative technology. I mean come on, cube storage?!?!
Too bad, they just can't make any inroads in the client side market. They invented the harddrive years ago and today they aren't going to even make any more client models.
Anyhow, I just wanted to talk about cube failures. Ice cube uses a 3x3x3 array of 27 cubes. But, the question is what happens if a cube goes bad. Essentially, you can never turn off Ice Cube. It's meant to be continuously running. If a single cube failure occurs the system just routes around it. To compensate you can stick more cubes on the outside. Of course, throughput will be hampered.
I asked the researcher what happens if say all the middle cubes burn out or when the throughput gets too damaged. He responds, "Well, given the failure rate, it probably won't be an issue until about ten years have passed, and by then we'll have much more powerful storage technology."
Finally, anything that is water-cooled is nifty in my book.
This article IceCubes would mean cool computing at New Scientist covered the technology.
"Software ... core challenge"??? (This sentiment is in the context that IBM aren't totally clueless about this sort of thing ;-)
Starting with a simple schema:
- Low level disk manager carves up disks into globally uniform chunks - say 20GB for argument's sake.
- RAID manager does the usual RAID 5 stuff using chunks from different cubes.
- Logical volume manager combines/carves up logical raid arrays into user required sizes.
- And finally a robust resizeable filesystem presents space to the user (or go back a step to present a virtual block device to Oracle or anything else that likes to avoid filesystems.
OK - that's a simple schema from which a better system can be evolved - but the core technology exists now. 1- disk partitioning; 2- RAID; 3-Linux LVM, Veritas Volume Manager and many others exist; 4- Growable filesystems exist (reiserfs, Veritas etc etc. Need to work on the ability to shrink for a fully rounded solution. Stage 2 needs to be careful concerning topology to avoid bad latency problems.To make this truely plug and play (but not in the MS sense) inserting a disk-cube would see it tested, auto partitioned and put in a pool. The systems engineer would be required to create/delete/alter filesystems and/or virtual disks as they needed - and configure things like how many simulatenouse cube failures can the system tolerate, how many hot spare cubes are kept in the pool and so on.
The software to do the underlying stuff is here today - I'm using it - albeit rather manually. The automation/management software to make this polished isn't hard conceptually. Of course if you only wanted one filesystem like the article mentioned it would require even less configuration ;-)
I'm actually much more impressed with the hardware here. Very cool. Not sure about the 3D and "stacking" structure. Bugger to replace a dead cube in the middle. Unless you are supposed to leave it there and throw a new cube on the top? I'd go for a 2D stacking system with overlapping layers (like a brick wall) - but with the couplers designed so you can knock a brick out sideways leaving the others undisturbed. Hmm - just a thought...
Why can't women be like Hedy Lamarr - beautiful, talented and inventors of frequency-hopping spread-spectrum techn
They seem to use smaller 2.5" hard drives, like the ones in notebooks. It does mean smaller power consumption and less noise, but what does that do to performance is yet to be seen. Maybe they are betting on time to make them faster and technologicaly more advanced. Yet, after i read an article at TomsHardware about doing raid with 2.5" disks, i am a believer! Not! :)
/Pedro
/Pedro
Dead, but reportedly alive...
Karma whorin' since 1999
I feel that the "lego" comparison is a bit flawed - this to me suggests a completly sealed box which stores data, power being inductively coupled, data through RF, etc. Also, lego is designed to be built, taken apart, built again.
This system is meant to have 27 cubes in a 3x3x3 cube, and when part fails, it is supposed to remain in place. Low latencies and high throughput are due to their being interconnected to the surrounding bricks.
First issue here is, that people don't like seeing things fail, and leaving them. This thing contains a "fast x86 processor", a gig of ram, (later on) six port Infiniband switches, plus all the disks. One of these failing is expensive - and getting the middle brick out would require removal of many other bricks, and probably knock out the system quite well....
It isn't really exandable either. For 27 cubes, perhaps the 3x3x3 is the best layout or topology of the blocks, but as you increase the size of the array (100 bricks or something), a cube becomes far more complex, with longer paths between cubes, longer latency, impossibility of removing a central brick. Heat would build up in the centre (yes, they are watercooled, but every part will be making heat, and not all of them connected to the heatpipe and watercooling system).
Maybe some mad buckyball style arrangement would provide the shortest average path between disks (but this would require a lot of statistical work, and depend on how the data was stored, what sort of access was required).
We could end up with huge, weirdly shaped storage arrays, like in films.
The watercooling is a step forwards, working in server rooms is getting far too loud.
Reliability may be an issue - 2.5" disks which it uses are known to be not as reliable as their larger counter parts. And there are a lot of them in this (12x27 = 324 disks), so failure is almost guaranteed within a short time.
I think this may be more of a concept thing than a final product - certainly the lego and modularity aspects need to be re-thought.
They're selling most of their hard drive business to Hitachi. They will continue to work with Hitachi on research, etc.
Hot Damn! It's the Soggy Bottom Boys!
I have to say, I wish I'd thought of it first.
However, this does bring me back to an original idea that I had for a server room. The room will be entirely empty, with large square tiles on the floor. Each tile will have information on the hardware that is below it (server name, switches, routers, etc). And, each will have a latch of some sort. Then, you unlock the latch, and pull up, and a large storage bin below slides up on spring-loaded rails, and locks into place. Then, you service the parts that need working on (swapping tapes, changing bad hdds), and slide it back down. All of the hardware will be sub-ground-level, which will make for much easier cooling, and a lot less cluttered environment.
These Ice Cubes from IBM would make a helpful addition to this idea, except you could only have probably two or three servers to a tile, attached one on top of the other. And there would be no side-to-side connections.
Eh, it was an idea.
The speed of time is one second per second.
read the article: you don't. You leave failed blocks in place and plug another one in on the top.
"The new wave is not value-added; it's garbage-subtracted" - Esther Dyson, Dec 1994
I'd imagine the software for this would work a lot like freenet, assuming it will be fault-tolerant and hot-swappable. Files would probably be scattered about in such a way that if a piece is temporarily unavailable, it would find the missing piece elsewhere, with the possibility of additional storage coming online or going offline randomly...
Please consider making an automatic monthly recurring donation to the EFF