I am building a Supercomputer, and need help!!

[Shalmanese]

lets say your base instability is 1. if you increases it by 1%, you have a base instability of 1.01 for each node which means the system would have a base instability of 1.01 ^ 36 which means you now have an instability of 1.43 or a 43 % increase.

At least, that is how I see it.

 

[ElFenix]

umm... beowulf isn't the only kind of cluster available... in fact... the only thing it really does is scale well. it scales really well. but it can't manage itself. and if a machine goes down you have to start again, no redundancy.

 

[Elledan]

<< umm... beowulf isn't the only kind of cluster available... in fact... the only thing it really does is scale well. it scales really well. but it can't manage itself. and if a machine goes down you have to start again, no redundancy. >>


That's why there is MOSIX

 

[Armitage]

<< umm... beowulf isn't the only kind of cluster available... in fact... the only thing it really does is scale well. it scales really well. but it can't manage itself. and if a machine goes down you have to start again, no redundancy. >>



Actually, that last bit is only true if the master node goes down. It's a b|tch, but you can write your software to tolerate the failure of remote processes & nodes. At least if you use PVM. I have had occasional trouble with PVM locking up if a node goes down, but thankfully that is a rare occurence on a good cluster. Failed processes are another story ... I've got a nasty heisenbug in my current project that kills slave processes on occasion , but my master process just detect the failure, and respawns the process.

 

[StattlichPassat]

"Could it be because of the crappy motherboard or PC133 RAM?"

Yes, and the SiS chipset too most likely- I would see if you can retest the AMD setup on at least an AMD 760 if not a VIA KT266A or nForce. Wait, is each system an SMP rig? Check into the new MPX chipset.

 

[Woodchuck2000]

I still find it hard to believe that someone is suggesting that we overclock a cluster. Overclocking is fine for single PC bedroom tweaker types, but to do anything serious? that's really asking for trouble. I'm just waiting for the same type to suggest overclocking a server...

Stability is paramount for that kinda thing. I would recommend using AMDs on an AMD chipset - They tend to be much more stable, if a little slower.

 

[Nefrodite]

but its such a big overclock would the decrease in uptime be made up for by increased speed? its not like a tiny overclock.

 

[Xen]

:|This thread is so bad, it made me want cry... so here is my first message on this forum.

1) if you have to ask somebody on how to make clusters then its probably shouldn't be you.
2) get a clear objective from the guys who is going to use it... I mean to the point of you getting hold of their program.
3) if you got the first two conditions out of the way, then you set down and do a cluster.

First few thing you don't do in a cluster

100Mbit lan (unless you like that "home made feel")
seperated stroage unit per node (unless you like to torture yourself)
use single CPU when the SMP is avaliable (unless you like being second class)
use AMD when Intel is avaliable (unless you made your own Intel cluster before)

Thing to do in a cluster

1)PCI Back Plane (Even at its lowest grade its 2x the speed of a 100 Lan, at its highest grade its faster then a 1000 Lan)
2)Single board Computers (There are many quad/duel PIII Xeon Board out there, Put 8 - 16 of them on a PCI Back Plane you just got your first cluster that still fits on your lunch table.)
3)Memory, Memory, Memory (RAM/Cache are worth their weight in Gold)

Everything else is base on what you want it to do, and how many time you want it.

 

[Armitage]

<< :|This thread is so bad, it made me want cry... so here is my first message on this forum. >>



Please enlighten us o great one

<< 1) if you have to ask somebody on how to make clusters then its probably shouldn't be you.
2) get a clear objective from the guys who is going to use it... I mean to the point of you getting hold of their program.
3) if you got the first two conditions out of the way, then you set down and do a cluster. >>



Ok, I agree with you so far, and have said as much myself.



<< First few thing you don't do in a cluster

100Mbit lan (unless you like that "home made feel") >>



And you can say this with no idea what the size of the cluster will be, and no clue what the application is?
For a small cluster and the right kind of applications 100MBit switched ethernet is plenty of network. And cheaper by far then anything else, although 1000baseT is getting there.



<< seperated stroage unit per node (unless you like to torture yourself)
use single CPU when the SMP is avaliable (unless you like being second class) >>



Again, you have no idea what the application is. I've worked with simulations that were extremely bus-limited that would run much faster on 2 single CPU boxes then a single SMP box. You've gotta do the benchmarks.



<< use AMD when Intel is avaliable (unless you made your own Intel cluster before) >>



Huh? This makes no sense. Why do you give your blessing to us peons to build AMD clusters only if we've built intel clusters before? Frankly, the $/performance ratio in floating point stuff is so much better then Intel that you'd have to be crazy not to consider it. As before though ... do your benchmarks.



<< Thing to do in a cluster

1)PCI Back Plane (Even at its lowest grade its 2x the speed of a 100 Lan, at its highest grade its faster then a 1000 Lan)
2)Single board Computers (There are many quad/duel PIII Xeon Board out there, Put 8 - 16 of them on a PCI Back Plane you just got your first cluster that still fits on your lunch table.) >>



I haven't ever used this stuff. How does the price compare to the more commodity level stuff you ussually see in clusters? And if you can only get PIII, you're way behind on floating point power. But again, if you need the bandwidth & latency that this can give you ... go for it.

 

[mste456]

sounds like u guys need a program called virtual machine. it installs it's self over the existing o/s . the clever thing is that it also alocates a certain ammount of sys resources to the software. i ur case the 36 pc's that are presumably networked would be linked as the one machine hense performance to die for

 

[overdoze]

WingZero94,
Ok let's concentrate on getting the task done shall we ?
I have build such a thing already and cost is number one issue, second is performance, and third reliability. Just imagine one node down would do what to you. Here is my detail of each box.

$215 Tyan Tiger MPX S2466N mobo build in 10/100 3coms NIC at newegg
$65 CHENBRO JASPER Series PC60220 with 300W PS at newegg or chieftech case with cheap 400W PS
$5 120 mm exhaust fan at KDcomputer
$123 80GB ide IBM hardrives 7200rpm at newegg
$123 second HD
$10 floppy drive at directron or frys
$22 LG 52X CDrom at newegg
$18 ATI OEM Rage "Mobility" VGA Card 8MB AGP at newegg
$21 Thermal Integration's Dr. Thermal TI-V77L heatsink at newegg
$21 second heatsink
$128 512M DDR ram at mushkin
$128 second memory stick (will work with 2 non register stick if use 3 or more have to buy register ram)
$150 AMD XP 1900 at newegg
$150 second cpu
-------------------------------------------------------------------------------------------------------------------------------------------
$1179 Total exclude shipping

This is roughly equivalent to 2 of your node each cost $600 and it is a kick ass configuration as far as stability and performance goes also save space power. If you want to add a 1Gb NIC it will cost you $200 more and I'm not sure you will need the second HD.
You might be asking or saying it will not work. Well I have it work in my company server room fine (rackmount however). I'm sure you can even find more bargain price if you buy 18 of these configuration at the same time.

Just my 2 cents

 

[Nitemare]

What exactly is backplaning? I found some pictures but have no idea what exactly it is and how it is implemented.

pictures

 

[davidsnot]

Here's some interesting reading. Might help a little.



Klat2

Friend of mine saw this running, it's fairly impressive. They talk about the interconnection topology as well..