| Click here to select a new forum. |
| G5 Supercluster |
Posted by: Quadraman on 2008-04-04 21:17:02
http://www.arc.vt.edu/arc/SystemX/index.php
A supercomputer made of 1100 G5 Xserves. [:O] ]'>
|
Posted by: coius on 2008-04-04 21:28:55
*yawn*
That was one of the main reasons the Dual 2.0's didn't ship immediately when Apple Released the G5. Virginia Tech took them all :-/
|
Posted by: LCGuy on 2008-04-04 23:39:56
Ayup. It was pretty extreme back in the day. IIRC they're using Xserves now, right?
|
Posted by: Aoresteen on 2008-04-05 00:33:54
What happens wen you need to upgrade the OS? Or defrag a drive?
Seems like a maintenace nightmare.
How many watts for power du they pull?
|
Posted by: ~tl on 2008-04-05 03:24:39
What happens wen you need to upgrade the OS? Or defrag a drive?
Seems like a maintenace nightmare.
How many watts for power du they pull? I believe Apple's server management software takes care of upgrades to machines remotely. You can just push a button and all the servers will update themselves. As for defragging, for the most part OS X takes care of that on the fly. How would it be any more of a nightmare to maintain than any other large cluster? I would be willing to bet that it would be a lot easier to maintain than most...
On the power front, the G5 Xserves drew about 400W. From the FAQ page on the VT site, it says that the whole cluster draws about 310kW per hour -- but I'm guessing that's for cooling, the networking hardware, etc, as well.
The main reason they built it was that it was cheap for the amount of processing power they were getting.
Still, this is VERY old news [😉] ]'>
|
Posted by: Bolle on 2008-04-05 08:07:12
Still, this is VERY old news [😉] ]'> yea... must have been 4 or 5 years ago when they did that with 1100 G5 towers...
|
Posted by: Aoresteen on 2008-04-05 11:40:37
That's why a Cray is more effeicient. You don't need 1100 CPUs to to get to the Super Computer range.
In any event, it's never the CPUs that are the bottlenecks, it's the I/O channels that always lag the CPU.
I'd be interested in seeing what type projects get approved for the CPU time.
|
Posted by: Christopher on 2008-04-05 13:01:09
I would get an Xserve just because its so much power in that tiny space. Plus it's rack mountable!
|
Posted by: Temetka on 2008-04-05 13:52:51
That's why a Cray is more effeicient. You don't need 1100 CPUs to to get to the Super Computer range.
In any event, it's never the CPUs that are the bottlenecks, it's the I/O channels that always lag the CPU.
I'd be interested in seeing what type projects get approved for the CPU time. Cray's are also very expensive.
Virgina Tech built the G5 Tower cluster for around 5.25 million IIRC.
They also used infiniband for networking. This provides for far more bandwidth and much lower latency than even GigE.
Also the whole point of the cluster was to prove that G5's and Apple hardware can be used for large scale scientific computing purposes. Something for which I think the team over at VT did an excellent job of doing.
|
Posted by: Bunsen on 2008-04-06 09:54:48
310kW per hour Watts are a time-based unit. kW per hour makes no sense
As I recall, grunt per dollar spent and grunt per watt of energy consumed put the VT cluster way out in front of any other cluster at the time. For $5m they got a $50m+ system.
|
Posted by: ~tl on 2008-04-06 11:05:40
310kW per hour Watts are a time-based unit. kW per hour makes no sense
Yeah, that struck me as a little strange too... but it's right there on the page [xx(] ]'>
I guess they just meant watts...
|
Posted by: tomlee59 on 2008-04-06 13:11:35
Or maybe that's the rate at which the power consumption increases... 🙂
|
Posted by: Cory5412 on 2008-04-06 15:25:00
yeah, when the VT cluster was created the Xserves were only on the G4 revision, so in order to get any notable amount of computing horsepower, they rounded up a bunch of then-fastest G5 towers.
As far as supercomputers go, I'd personally prefer something like an SGI Onyx, capable of scaling to 512 or 1024 processors, and it's all on a single image, uses shared ram and cache, etc. Managing it, if you don't have to integrate it into an LDAP environment, is just as easy as managing one desktop computer.
Plus, they (nekochan.net) have ported a bunch of open source stuff, like firefox/thunderbird et al. to IRIX, so it's almost like using a modern desktop computer in that respect.
|
Posted by: paws on 2008-04-07 00:45:02
Firefox on 1024 processors?
It might be almost usable...
|
Posted by: iMac600 on 2008-04-07 03:21:35
Firefox on 1024 processors?
It might be almost usable... :lol:
Awesome response.
|
Posted by: Cory5412 on 2008-04-07 07:59:12
So true, so true. Maybe we can try to coax flash into running on it.
|
Posted by: Quadraman on 2008-04-07 08:38:20
Firefox on 1024 processors?
It might be almost usable... Nah, I think it'd still find a way to run your CPU usage up to 99%
|
| 1 |