http://www.spacemart.com/reports/Maryland_Professor_Creates_Desktop_Supercomputer_P rototype_999.html


Looks promising. The prof working on this has been at it since 79' so I have a good feeling we may see some of his patents filter into our systems in the near future.


Cheers,

Wonder how much power the air conditioner takes to keep it running cool. :p

Looks pretty interestng, and I am not at all surprised to see Defense involved in the funding. Smart move on their part. The Defense involvement may also limit how much we see filter down to personal use, at least in the short term.

Sounds like programming for it will be reasonable... however Microsoft / Apple / Linux OS's will have to be re-written in order to take advantage of what it would be capable of. We're talking a rebuild from the floor up, even more extensive than the jump to 64 bit.

Then all of the software would have to be written for it. Don't expect to be able to build a parellel core box and have your favorite games run on it...

Then again, in theory they should be able to simulate serial processing to be able to run "legacy" programs.

Sadly, though, our bottleneck is still our storage mediums. Crystaline storage, where art thou?

I personally think the software is the bottleneck 64bit architecture has been around for quite some time but the software side of things is very lacking.

But I would think the next bottleneck will be cooling and power as Byl slightly mentions. The faster things run the more they consume and the harder they are to keep cool.

Out of curiosity what makes you think storage is the bottleneck? I guess having a few DS4800s and DS8000s here running 15k drives and not being able to cause noticeable performance hits on the storage makes me think otherwise. If anything our performance hits come from poorly written apps that are riddled with memory leaks =\

Well, lets say you're running a system with PC3 11000 DDR3 (1375Mhz) ram. Each ram module is capable of 11GB/s. In dual channel configuration, thats 22GB/s throughput. Not too shabby.

Now lets go to down to communication between your hard drive and your system bus. SATA2 is 3GB/s. When SATA3 starts being used that'll be 6GB/s. With raid you can raise those potential speeds. So, 6+GB/s... still not too bad.

Now the hard drives... A spinning platter hard drive, even at 15,000 RPM, has a maximum burst throughput of around 125MB/s. Even raided, that doesn't NEARLY approach the FSB speeds in even a common computer.

My gaming rig at home uses DDR2 800, PC6400 dual channel ram. That means I have a theoretical throughput of 12.8GB/s. Running a throughput test on my raided SATA2 hard drives I got a max burst of ~95MB/s.

For the time being, the best speed you can get is from RAM hard drives. When a SATA3 RAM hard drive set up comes out and is reasonably priced, I may look into it. But in order to set one up, you have to buy the device itself, plus 4-8+ ram modules. Even with 8 1 GB sticks and with cheap ram (you're going to hit your 6GB/s bottleneck even with the cheapest DDR ram) thats still only an 8 GB hard drive. You can put your OS on there, but you'll still have to use another drive for storage.

Crystaline hard drives are supposed to be under development but you never hear much about them. Supposedly in theory you can use special crystals for storing mass amounts of information and use intersecting lasers with no or very few moving parts to read and write. I learned about it while I was in school, but there isn't a whole lot of information out there on it.

Either way... we need something faster... and no moving parts seems to be the way to go.

Thank you for the post and explanation. I have to be 100% honest with you I've never thought of storage in such a way... namely b/c in my dealings with expensive SANs I have yet to see the storage experience bottlenecks and instead see the issues on the application side of the house.

I've never heard of crystaline storage either; something for me to look up during slow days at work :)

Now regarding speed of drives did you consider the cache that most SANs have in your throughput numbers?

Well, when reading / writing off off multiple hard drives in a large enough array you can start approaching the point where you're actually going to be bottlenecked by your communication medium. However, that doesn't change the fact that even the fastest hard drive is still the slowest point in the system

As to caches, they are pretty much a temporary ram hard drive. Fast, sure, but you only get that benefit when you are accessing stuff that still resides in the cache. What happens when you try to pull up a large file that hasn't been accessed in months? It still has to read off the 110-125MB/s hard drives.

With a large enough cache and if you're constantly accessing the same files, you will see great speeds from something like that. But that is more of a combined solution to the slow storage problem. I know a company that paid a LOT of money to have a RAM storage server built. Its fast as hell, but spending a ton of money to try to approach the throughput speeds of our system busses should not be the only option, and sadly thats what we're faced with.

Oh... also of note: If you're going to have a RAM storage unit built you better have a fiber backbone between that server and anything accessing it, otherwise you're going to have a network bottleneck, even with Gigabit (~125MB/s).

Well, when reading / writing off off multiple hard drives in a large enough array you can start approaching the point where you're actually going to be bottlenecked by your communication medium. However, that doesn't change the fact that even the fastest hard drive is still the slowest point in the system

I don't usually do this but it's easier to keep track of each point :) Very true and in fact in the envioronments I work in that is the only "hardware" type place I have seen bottlenecks. In fact I am currently researching how to install TPC for storage in our data center. We've been using other tools and IBM keeps telling us to use this so... Nonetheless I just got 4 new 4gig fibre adapters that will be used for our new DS4800s (4gb) I'm excited to see how they perform since IBM basically scrapped the DS6800 in favor of the cheaper 4800 b/c it's kicking big brother's butt in performance and can have more shelves and 700GB drives. Anyway a bit off topic but so far it's the only place I've seen bottlenecks granted my tools are limited but hopefully changing.

As to caches, they are pretty much a temporary ram hard drive. Fast, sure, but you only get that benefit when you are accessing stuff that still resides in the cache. What happens when you try to pull up a large file that hasn't been accessed in months? It still has to read off the 110-125MB/s hard drives.

Well if you have old crap that doesn't get accessed much you certainly don't put it on your tier 1 storage you instead put it on cheaper slower storage. Being a hospital we do this for PACs (picture archive computer system) which is basically x-rays. Every image taken at our facilities is digitized, cataloged and stored FOREVER. As it ages it moves through the organizations many archives until it eventually winds up on the slowest cheapest one that is currently several TBs though it is still replicated back through the other archives in case we ever lost one. Anywho our heaviest hit databases/applications are all on high end storage with max cache per controller. The nightmare in being in healthcare is that you never purge and someone has to figure out DR and backups for this heap of data. We tier our backups like we do storage but we're still pushing 2TB+ a night in data to tape to be sent offsite. We just bought two IBM directors 256B ones for our future growth and addition of SVC to help manage all this and replicate from site to site.

With a large enough cache and if you're constantly accessing the same files, you will see great speeds from something like that. But that is more of a combined solution to the slow storage problem. I know a company that paid a LOT of money to have a RAM storage server built. Its fast as hell, but spending a ton of money to try to approach the throughput speeds of our system busses should not be the only option, and sadly thats what we're faced with.

Thankfully we're not there yet. Our network group is still trying to catch up to our data center :)

Oh... also of note: If you're going to have a RAM storage unit built you better have a fiber backbone between that server and anything accessing it, otherwise you're going to have a network bottleneck, even with Gigabit (~125MB/s).

We have 2 hospitals and 80 some clinics between Northern Maryland and South Central PA. The main hospital, data center complex, and DR site are in a fibre triangle. We run 3 T3s to Gettysburg hospital and a couple of T1s for backup. To the larger clinics we run T3s, smaller ones get T1s and god knows what else (someone elses headache thankfully) Verizon loves us it's one of our larger expenses. Still 9 times out of 10 our issues are either network or app related. And the only time we had "hot disks" on our SAN was b/c the database was configured incorrectly at the OS side. Using AIX and Logical Volume Manager one can manage disks very efficiently not to mention the way the DS6800, DS8000, and DS4800 are designed (allot like the LVM) so it is "supposed" to be evenly distributed. We have our most important database spinning on 96 disks and when a write happens it spins all of those disks not one or 8. It was a total PITA to set it all up but it works great. Now if only IBM will fix their LVM bug with HACMP we'd be set but that is a different topic. :D





Do you happen to have any experience with IBM's Total Storage Productivy Center or line of products? Know of anything else that helps monitor SAN performance? We are mostly a Computer Associate's NSM shop for enterprise monitoring and are testing their stuff but we're always looking for other ways to keep an eye on what's going on.

To be completely honest, you have far more experience with major network applications than I do. I'm still pretty fresh out of school (2 years) and haven't worked on anything major yet. I'm getting good administrative experience at the company I'm at and eventually I'll move on to something bigger.

So to answer your question, no, I don't have any experience with TPC or any major SAN management.

You still make a good point. How many desktop computers run fibre channel SAN arrays for storage?

This isn't about a massively parallel super computer in a data center packed with top tier storage configurations. Those are old news anyway. This is about something to sit on your desktop, where storage/memory bandwidth limitations likely are still the most prominent concern.

You're right Sanchek but our conversation moved away from that at least I wasn't trying compare this desktop supercomputer with SAN storage. I'm currently trying to get a handle on our Storage infrastructure and was hoping to get a conversation going with someone who appeared to be very knowledigible on the subject so do forgive my derail if you deemed it inappropriate.

Akom while you may not have allot of experience your thoughts are still interesting and I tried to pry your brain a little more and I enjoyed the conversation nonetheless :) I'll drop you a PM instead of derailing the thread anymore.