Linux cluster help

[pl8er]

I wasn't exactly sure where to put this topic so this is where I decided. I'm going to be trying to build my first linux cluster and would like to know what everyone thinks would be the best software to load on this setup. As of now I will have my master node be a 1.8 P4 with 512mb ram and 20 gb HD. I have 3 slave nodes that are 933mhz P3's with 128 ram and 10 gb hard drives. Anyone have suggestions on what flavor of linux to install and what would be the best cluster software to use?

I appreciate any help.

Justin

 

[drag]

Well there are a few different types of 'clusters'.


You have load balancing clusters, high aviability clusters, high performance clusters etc etc.

For the easiest way to get a cluster going is to check out http://openmosix.sourceforge.net/

They have boot cds and such that will turn any network of 2 or more computers into a load balancing cluster.

It's a good place to start.

 

[pl8er]

Now by load balancing you mean that all the nodes will get information from the master and do the work? Is this the same as making the the computer think that its all one big multiprocessing computer? Sorry if these are basic questions, I'm still at the start of all of this.

Thanks again,

Justin

 

[drag]

What it means, basicly, is that any program will not run faster then any single machine can run it... However if you run many programs they will load balance between 2 or more computers for maximum effect.


So for normal desktop stuff it's not going to help (since any thing tied to high I/O or to the display will perform worse after it's migrated if thats possible at all). It doesn't help with multithreaded applications since there is no universal memory namespace...

But it's good if you have to do things like do batch rendering, multiple transcoding or mathmatics that can easily be split up into multiple sections and run at the same time on different programs. that way the load balancing mosix style system will actively migrate the programs from system to system based on aviable resources and on a algorythm based on how much benifit vs loss it is worth to move a proccess over a network.

It's fairly complex, but OpenMosix is about the simpliest system to get running. Also it doesn't require special programming skills to develop programs that will run on it, like most High Performance Computing (HPC) stuff like what is used on scientific computing (think Fortran and MPI libraries) and Beowolf-style nodes. Any regular program will work with it, with some exceptions.

 

[pl8er]

Thanks for the links! Also thanks for the information. Now is there any setup that actually makes all the computers share work? I'm thinking about running something like Seti@home on it just to watch it work..is there a way to only run one file at a time and have all 4 machines working on the project? I've noticed that there are all of these different types of clusters such as SMP? and many others. I would like to make the master tell the slaves what to do and when they return the data have it give them more to do and the master actually give the end result. Does that make any sense? Is that even possible. I'm totally a newbie at this and just like one of the sites I read I guess I'm a drone bee in the beehive just trying to taste the honey

Thanks again for the input, I really appreciate it!

Justin

 

[erwos]

Sure, openMOSIX will do work sharing. However, I guarantee that you'll find it utterly unusable for 99% applications, because threads that need memory sharing are going to rip apart performance. Memory accesses over ethernet - yes, that sounds like a winner.

The whole "master tells the slaves what to do" is what libraries like MPI are for. It's not something that just happens because you wish it to.

You don't need a cluster. That much is apparent to me, at least.

-Erwos

 

[pl8er]

I honestly don't recall saying that I needed a cluster and I don't recall saying once that I knew what I was doing. I'm asking for input on a project I'm trying to do. I want to learn linux and I want to learn basic clustering, from there I may try for more advanced clustering.

I know that for information on something this advanced that this is the forum to go to. 99% of the people here want to give advice and then you have the 1% like yourself who self proclaim themselves the worlds smartest computer person and want to bash on everyone for admitting they don't understand things. Why don't you go back to building your cantennas and stealing your neighbors internet and trying to figure out how to get a date and let me try to learn something that I'm interested in.

Justin

 

[Armitage]

Reminds me of how I got started with clusters - 5 486's @ $15 each and a couple of weekends to figure it out

I would suggest whatever linux OS you are most comfortable with. At your current level, the differences aren't going to matter much unless you want something like Mosix as described by some other folks above.

But the question comes down to software? There's some stuff available that works on traditional master/slave style beowulf clusters such as you describe - PVM_POVRay comes to mind. There's probably something for SETI as well, though for SETI I don't see how a cluster type config is going to really help you. Almost all of this kind of work is scientific type programming internal to various research groups, etc. and not available (or of interest) to the general public.

The fun part is rolling your own to solve a problem of interest using something like the PVM or MPI libraries to handle the message passing between boxes. IME PVM is much easier to get started with and much more robust (though I've been told that the latest MPI addresses alot of this). But PVM is old and not developed anymore, which shouldn't be a concern at your level, but may be a dead end down the line.

Any more specific questions?

 

[pl8er]

I appreciate your input on this! Thanks for the information. I am thinking about using either red hat or ubuntu (spelling?) on my master and using ParallelKnoppix on the slaves for a diskless setup in the beginning. Just to get used to running the system then possibly installing an OS on each node. I really have no specific use in mind as far as programs go. Also, I am not a programmer. I guess this is a do it to try it out thing. What did you run on your first or even current cluster?

Thanks again,

Justin

 

[Armitage]

Originally posted by: pl8er
I appreciate your input on this! Thanks for the information. I am thinking about using either red hat or ubuntu (spelling?) on my master and using ParallelKnoppix on the slaves for a diskless setup in the beginning. Just to get used to running the system then possibly installing an OS on each node. I really have no specific use in mind as far as programs go. Also, I am not a programmer. I guess this is a do it to try it out thing. What did you run on your first or even current cluster?

Thanks again,

Justin

I'm more interested in clusters from an end user/programmer standpoint, so the first thing I ran on it was likely some cluster version of "Hello World" - you know, simple C program that dumps back "Hello from Node 1, Hello from Node 2, etc."

The first serious thing I ran on it was a parralelized version of a genetic algorithm I had developed to find optimal satellite constellations. Stuff since then has been of the same flavor - space mission optimization, sensor coverage analysis, collision avoidance, etc.

If your not a programmer, I'm not sure how interesting this exercise will turn out to be for you. You'll just have a network with some somewhat odd configuration choices. You might look for that PVMPOVRay program I mentioned - if you're interested in ray tracing, that might be useful. Haven't tried it myself.

 

[drag]

Any task you want to do that can be parrellelized is usefull for clustering stuff..

Raytracing is a great example, although to work with OpenMosix stuff they need to fork() the task into multiple proccesses rather then run multiple threads. Yafray should be able to do this, as should Povray.

For some apps, like Maya or Cinelerra, for instance, have their own 'node client' applications for doing clustering applications.

Stuff like that.

For 'real' computational clusters you use libraries developed for clustering and maybe OpenMosix a bit to reduce administration overhead by load balancing clusters.

Then there are different types of clusters for doing other stuff. Like if you have a lot of file servers you can use clustering stuff to make sure that they have high performance and strong failover. This is what Google uses clusters for.. They have 20 machine clusters, or something like that.. as a single node.. then they'll have a few dozen of those. Each machine load balances within it's own cluster node and shares the load, databases and I/O, and such, then they cluster of clusters that are used for further load balancing and also for high aviability.

So if you have a machine have a hardware failure on one PC, that entire node goes down automaticly and a spare node kicks in and the effect is completely transparent/automatic. Then techs can come in and repair the faulty PC on their leasure with not even a millisecond of downtime in the eyes of a end user. That way google can just use a small handfull of experianced techs do the same work that it would take a army of Windows techs if they wanted to try to do the same thing with traditional servers.

Also another thing to check out to see what people will be using more and more in the next few years check out Xen.

Xen is a machine emulator that is very fast and is much like the 'server' version of Vmware. Instead of running a Xen program in a host OS to run another OS emulated Xen runs by itself and hosts slightly-modified OSes on itself.

It can do some very special things like emulate with only a 3-7% in reduction over performance on straight hardware and in a cluster situation it has the ability to actually migrate OSes from machine to machine WHILE they are running!

For a experiment Xen guys hosted a little Quake3 LAN tournement running with it's server on Linux running on Xen. As you know latency to responses is very important. During the middle of the game they moved the running server image from one PC server to another. The end users didn't even notice the downtime.. they calculated the system downtime in the transfer over at about 30msecs or so.

So that's something that is very cool.

 

[pl8er]

Wow! Those are some really good ideas for things for me to try. Thanks again for the ideas and I am sure I will try most of those out! I'm sure most of this project will be install...mess around...format, install something else...etc etc...I want to get a working knowledge of how this all works and try as many different versions as I can.

Thanks again for all the input!

Justin

 

[drag]

That's one of the cool things about Xen.

You can easily setup multiple OSes to mess around with and load and start them up with Xen. I haven't had much of a chance to play around with it myself, but I expect that you could do something like...

Setup you main machine with a Desktop-style setup with X and everything. Run that in 'ring 0' on Xen so you can get access to hardware for things like X (I don't know how well third party drivers like 'nvidia' would work out though) on you main node.

Then setup you client nodes with a minimal install of Linux running on Xen themselves.

On the master system you have a fairly large harddrive, like 120 gigs or so. Or a raid array, whatever you want. On top of that you have LVM installed so that you can have many little 'partitions' (logical volumes in LVM terms). On little 1-3 gig partition you can keep a minimal install of some Linux version.

So then you can uninistall, format, and keep many different versions of different clustering versions with different installs and different setups without having to format you main systems and your nodes all the time. Plus with Xen your abstracting from the hardware so that you don't have to mess around with hardware configuration all the time.

So you can do things like 'oh, today I'll run a OpenMosix-based cluster'.. and then tommorrow run a OpenSSI cluster, then the next day run a OpenMosix at the same time as a OpenSSI cluster. Then after you get bored with that you can run a Kerrighed system so that you can migrate threads. All without having to reformat anything.

 

[pl8er]

That honestly just blew me away LOL....I will look into that for sure but I think its a tad beyond me for the stage I'm in right now. I will look into Xen for sure though, it sounds very awesome.

J

 

[drag]

Ya it blew me away too when I first started reading about Xen. I messed around with clustering a bit mysefl and it was a huge pain because I'd make the changes to my system to work in a clustering setup and when I would want to change I'd have to format and reinstall, which is a always a huge pain.

But Xen should make things much much easier. Hell, to play around with clusters you can even do it virtually on one computer pretty much.

It's all very new though and I'm working on other things at the moment.

I'm looking forward to when the next gen of AMD and Intel proccessors come out.. Intel calls it 'VT' technology.. but AMD should have similar stuff. It'll add some extra features to hardware-level memory manager to make running things like Xen more efficient/easier, also should make it possible to run systems with less porting effort. For instance now it's only possible to run Linux with a modded kernel and a couple different versions of *BSD in Xen.. with the VT stuff it should make it easier to run other OSes. (probably including OS X and Windows).

Still though, it's very new stuff.

 

[MadOni0n]

what kind of programs benefit from clusters? can i gain any benefit from clusters from normal progs or its only if i have mathematical data to process?

 

[MadOni0n]

what kind of programs benefit from clusters? can i gain any benefit from clusters from normal progs or its only if i have mathematical data to process?

 

[kamper]

Some guys in the Distributed Computing forum set up a Beowulf cluster to cruch seti classic a few years ago and I believe it's still running. I bet if you pop your head in there, someone will be happy to dig up the thread and answer questions. Maybe not what you want to end up doing, but worth looking at...