home network cluster project

metalh34d · Tux's lil' helper Joined: 04 Feb 2004 Posts: 126

I have a collection of servers that I keep in my closet. They're all older model pcs with outdated hardware and small/slow hard drives. I figured the best thing to do with them is to cluster them all together and create one super server for my personal use. An extra benefit I saw would be if I could have my desktop tap into the added power of the cluster when I am compiling/running proccessing power hungry apps/gaming.

My issue is I have very little knowledge of how to setup such a cluster. Learning how to do it isn't even the real issue here. What I really need help with is determining what I setup I should use to get the job done the best way I could. The term "beowulf cluster" is what attracted my attention to creating my own cluster years ago. Looking through the gentoo forums (gentoo being the distro of choice for the cluster) I have ran into a lot on mosix, openmosix, and openssi. Adding beowulf cluster into that list I now have 4 different setups I could use and little knowledge on all of them. What exactly is all this technology and how does it benefit me?

bunsen · Tux's lil' helper Joined: 10 Aug 2003 Posts: 105

Clusters serve varied purposes. Some clusters are set up to provide high reliability, with hot-standby.

The mosix and open mosix cluster are for balancing load. A job running on a heavily loaded machine can migrate to another node in the cluster, transparently to the software and the user. This is not parallel computing though, as the entire process migrates. If the big job can be broken down into separate ind independent tasks, then there's a chance they'll be shared out around the cluster. Not all jobs can migrate. It depends on what system resources are involved.

Beowulf is commonly associated with parallel processing. The software must be written to do work that way. The cluster for such software is often built for that purpose, rather than a general purpose parallel cluster.

Condor is another cluster option. It requires that jobs can be run in batch mode. A scheduler delegates the jobs to nodes on the cluster.
It's conceivable to have a cluster of clusters. One job being given to a one cluster or another, depending on load.

There are not a lot of desktop apps that make use of parallelism, since the software must also be suitable for the cluster. As that's not always known, such software is probably more likely to be found in research. Not always, but that's what I've read.

Taking things further, not all tasks will benefit from parallelism. Some things must run serially. An iterative analysis process that uses the results of the previous calculation as the start of the next is an example. However, if the same analysis was to be repeated for different input conditions, then multiple analyses could be run at the same time, each using a different set of input data.

For some jobs, the act of migrating or splitting could cost more than keeping the job in one place.

Does this help?

metalh34d · Tux's lil' helper Joined: 04 Feb 2004 Posts: 126

that does help a lot. Right now I guess what I'll want is to continue my original plans for a distcc network. Do you think I can use some form of Mosix to run a dedicated server and still use distcc on the nodes?

dannycool · Posted: Thu Aug 26, 2004 5:26 pm Post subject:

If you're planning to use distcc, your best bet is to set up a nfs server and have the other machines boot from it. This way you'll only need to manage one portage tree

And it's some sort of a cluster. While no processes are distributed or anything, everything you do on the cluster you can do on any node you want, with full transparency.

metalh34d · Tux's lil' helper Joined: 04 Feb 2004 Posts: 126

That sounds great but I was looking into running a dedicated server which would use up a lot of proccessing power which my machines don't really have. I figured them working together would make a pretty beefy solution though. The added benefit of distcc and one portage tree would be awesome for my desktop and laptop though.

metalh34d · Tux's lil' helper Joined: 04 Feb 2004 Posts: 126

ok right now to my understanding openmosix will provide the load balancing for a server that expects multiple clients logging in and accessing information from a database.

There were some threads in this forum that were talking about openmosix also doing the same effect of distcc, sharing the compilation of an ebuild. Distcc comes off as the obvious choice for compiling ebuilds, but wanting to get more it sounds like openMosix is the solution for that.

It doesn't sound like setting up a performance cluster works for personal gaming purposes. If this is untrue, correct me please. However if I wanted to setup a dedicated gaming server for a lan, would I benefit more from a performance cluster or a balancing one. Does the performance cluster even work for this case? Does the balancing cluster even benefit a dedicate gaming server?

bunsen · Tux's lil' helper Joined: 10 Aug 2003 Posts: 105

With openmosix, watch out when building software, maybe specifically ebuilds. Some ebuilds would bomb out and I worked out it was something to do with process migration. An example was xfree I think. As soon as I stopped openmosix, it went OK. It's probably not necessary to stop it in that case, instead just tell it not to allow migrtion for that process. However, child processes might get overlooked.

On this basis, distcc is probably a better bet, since the end result is the same as what you'd like from openmosix.

metalh34d · Tux's lil' helper Joined: 04 Feb 2004 Posts: 126

however distcc doesn't offer the feature of balancing the load that I would ideally setup for my home networking projects; ie: dhcp, remote access, intranet site, local file transfer, e-mail experiments, lan dedicated gaming. These are all projects from the near future I will be working on and would ideally be the main focus of utilizing the extra computers I have lying around. Clustering seems like it would end up being the best solution since proccesses would be handled dynamically depending on what needs to be accomplished. Distcc for my desktop, laptop, and updating the nfs server would be a huge benefit, however the plan on how to utilize that is unsure of. Right now it seems like openMoxis could accomplish all of this and distcc's objectives (if slower) but then I'd have to run a moxis kernel on my personal computers which are very happy running the latest nitro sources (go Con!) right now.

sPudlE · n00b Joined: 28 Aug 2004 Posts: 4 Location: dallas,tx

what type of cluster would be suggested for running setiathome? or gipsS?

-sean
welchs@teov.org

bunsen · Tux's lil' helper Joined: 10 Aug 2003 Posts: 105

Apparently the setiathome client does not benefit from parallelism, so you can probably rule out parallel processing clusters.

My only direct experience with clusters has been using openmosix, so that's what I'll limit my comments to.
Supposing seti would migrate successfully on an openmosix cluster of N nodes, and you other jobs forced one seti process onto another node already busy with seti. That node is going to now be running two seti processes at half of the earlier speed.
If migration is OK and you install it on one node but run N seti processes simultaneously, you need to be careful with the seti configuration to avoid the results/status files of all processes colliding. Different working directories are needed for each process.

It might be simpler to just install it onto any machine that you want it to run on and run one seti process per machine.
Unless your value of N is big enough to justify a script to deal with setting up different working directories.