Printable View
Can enyone explain to me and to others, something about Clusters?
What's the Clusters types?
How it works?
and eg.
-----------
I'm interesting for Cluster for SETI@home:
but no BeoWolf cluster, I need cluster when I can use all computers as one fast maschine (not meny 'slow's). I need onemaschine with power of all...
If you understand?
Do you mean a farm of PCs linked in a network or do you mean multi-CPU machines?
I know how to build a SETI Farm, and BeoWolf (Linux) Cluster (a farm of PCs linked
in a network).
For this you need few PC's (Clients) as NODes, and one HEADnode (Server).
Every NODE running SETI Client separately:
eg. If you have 5 PC's with AMD AthlonXP CPU's and every of them need 2,5h for WU.
For 2,5 hours of computeing you will have compleated 5 WU's.
This is SETI Farm (build as BeoWolf Cluster).
---------------
And "multi-CPU machines" are something diferent... (Multi CPU OPTERONS... or Athlon MP...)
---------------
Cluster is a collection of interconnected computers working together as a single system.
I need some like HPVM (NT Clusters).
I want to connect few PC's to (G)LAN (Gigabyte LocalAreaNetwork).
When I run some application on HEAD_NODE it supouse to share CPU/Memory to NODES...
...so if I run SETI (for eg.) on HEAD_NODE it will compute one WU for short time:
Time to compute WU on HEAD_NODE = Time for compute WU on one NODE / Number of NODEs.
so if you have 5 HODEs time for one WU would be about 0.5h (30min) (becouse 2,5h/5nodes = 0.5 h)
Understood?
I need some cluster who will use NODES as one (Single) Computer.
Is this possible?
What cluster type I can use for this?
What OS need to be running on NODES/HEAD_NODE? Windows/Linux (FreeBSD/NetBSD)?
I'm not an expert, not at all, but i guess using Linux and with the right configuration you could do something like that, assigning a folder to each computer and then assign cooperative work, all the cpu's woking on the same WU, insted of one WU for each cpu. We usted to have a member here that could help you, and if i'm not wrong he had posted some info here in our forum about that matter, just try to seach a bit here, it may have some time by now...
This is a node:
JPEG Image, 1200x1600
http://us-distributed.com/images/node3/node31.jpg
Source: US-Distributed Team
Good one :shock:
Check this one out Here
Interesting...
:shock: :shock: :shock: :shock: :shock: AmazingQuote:
Originally Posted by TA_Super_10
OK
All this amazing Computers, are 'just' SETI FARM's (SETI STACKS).
Enyone have some information about some other Clusters?
Not 'just' SETI...
here are some Linux Amd Clusters
AMD Clusters
Or fancy having one of these AMD Clusters here
I have a question: Do you want a true multi-processor SMP or an ASMP configuration. I'm asking because with SETI and a lot of DC projects going to BOINC... they sniff out / detect farms. There is discussion about preventing farms from being allowed. (I was in boinc beta until recently).
There are configurations you can build (like SGI "numalink")... this is where multiple SMP machines are linked together (typically blocks of 4) to form a single machine. Only 1 has an ethernet port & AGP enabled / installed. These machines give a fairly low latency across the link, functioning still as one machine.. The next tier is to link multiple 'numalink' machines together. This is equivalent to building something like an Origin 3000 (which is a 'scaleable supercomputer'). You then 2nd tier these machines. In all configurations, it still looks like one computer... but has a buttload of processors.
If you permit me to refer back to the VERY old days of the VME bus and Motorola 68030 processors, We would put up to 16 processor boards in a rack on one backplane. Each board was linked to the other on the 'front side' to form a single machine for near SMP performance. We could then link multiple backplanes to make it as big as we wanted and had power for.
Given that the Athlon64 (Clawhammer) and Opteron (Sledgehammer) are basically the VME on a chip (first step of a multi-step process), taking the entire product line forward toward multi-core CPUs , soon to be followed by SMP multi-core machines, and then these will be linkable at the backplane/bus level that use 1024 bit wide memory busses, etc etc.
Am I catching what you want to build or have I missed the boat entirely or are you looking to build a 'node' that is heterogeneous and/or loosely coupled and essentially 'invisible' because it is behind a single IP address and the lead machine farms out the work to the other processors and handles all the reporting back to the project?
I do apologize, the hour is late here.