Creating a Windows Based Cluster

Ok I got this idea from NeuTron. Feel free to critize...

Ok I have a windows based netowrk that I run for a school, around 45+ computers. I was wondering if it is possible to create a cluster style super computer with windows 2000, NT based? They are all varying types with different hardware. is this possible or am I just dreaming again. It would have to be an underlying program that wouldn't bother other programs in use, so teachers can still input grades and check the web.

I would like to have it all be run off of one computer, that is linked in via a VPN via the web, as the control center... Mainly to run complex program compiling and such...

Give me thoughts guys....

-Shugart

Have you researched what Microsoft TechNet has to offer for help?

nah, but I sure will...

thanks
Shugart

OK, that site answers any and all of my questions... thanks

Shugart

Care to tell us your discovery?

Ahh, sorry, bad habit...

From that page I found out that to do a windows based cluster on a large scale would require me to upgrade to windows server 2003. Which is made to be able to cluster quite smoothly from my understanding. for me it is to much of a hassel that will come out of my pocket book, hihi.


But if you do decide to go with server 2003, it has built in remote clusting ability also. so you could in theory ( as I understand it ) have an internet based VPN and run up to 100 computers clusterd. no more than 100, though, I found that a bit odd, but oh well. a very interesting program none-the-less. but not worth the money...

-Shugart

You could do that with windows 2000 Advanced server as well, but be warned it sucks ass. Other than that you would have to build cluster software instead of microsofts canned packaged version. There are several companies outside of microsoft that will accomplish this or you can tap into Microsoft cluster APIs through NT on up via prgramming languages (IBM, Computer Associates, visual basic, visual C, etc). There was some rudimentary cluster capability with a service add on to NT. But these clusters don't really act like distributed computing where you steal a little bit of CPU power from a large base of Computers. They are totally dedicated to the cluster.

Also using MS clusters you will run into limitations such as shared data sources etc. I have a test Active/Passive behemoth cluster running an MS Exchange environ in 2000 Advanced Server. It was failing constantly until about 5 months ago. Supposedly 2003 is 100 times better.

Finally I have been monitoring some open source projects, too bad you don't have Linux :D
OSCAR is promising but like windows it's an application interface used to write your own stuff.
Oooh just thought of it, you can also look at bittorrent and use it's distributed capabilities to adapt you your own needs and it's win32 compliant.

\EDIT forgot openmosix - I listed linux projects because they can me adapted to win32 and Server 2003 on all your PC's is not really feasible.

Quote:

---

You could do that with windows 2000 Advanced server as well, but be warned it sucks ass. Other than that you would have to build cluster software instead of microsofts canned packaged version.

---

MS Clustering will not work as this person has defined their question. MS clustering is for availability, and the version that comes with Win2k does not suck in my opinion. It is one of the best clustering solutions available for any operating system, but it does not give you native CPU sharing for all applications. Win'03 clustering just gives you the ability to go up to 8node clusters without having to buy the datacenter version of the software.

If you want to do a shared CPU cluster you would have to use something other than MS clustering. MS Clustering just gives you the backbone to be able to run cluster aware software. If the software itself is not cluster aware the only thing you can do is run the non cluster aware software as a resource, when the resource fails on one machine, you can set actions that will be performed to start it back up. It will not allow software such as a graphical compiler to perform CPU sharing unless the software has the CPU sharing algorithms built into it.

Even exchange2003 is not doing CPU sharing. It runs all of its resources on an active node, if that active node fails for some reason it will either go to a spare passive node, or double up resources on another active machine. SQL server however, does have the ability to break up different tasks to different cluster resources and "share" CPU in that fashion. One of the only pieces of corporate level software that I am aware of that does it very well.

Quote:

---

Also using MS clusters you will run into limitations such as shared data sources etc. I have a test Active/Passive behemoth cluster running an MS Exchange environ in 2000 Advanced Server. It was failing constantly until about 5 months ago. Supposedly 2003 is 100 times better.

---

No offense but it had to be a configuration issue. Clustering is very specific about firmware levels on all devices, driver levels, and device configuration, not to mention that there are a lot of cluster specific hotfixes pre-SP4 that you must install to fix clusdisk.sys issues. I have 10 clusters in production each with over a terabyte of data in shared SAN diskspace and we experience phenominal uptime. I've not had a problem that was related specifically to clustering in over 3 years. Only 1 outage over the last two months, and that was related to a PCI device failure that crashed the entire server instead of just failing over to it's redundant device. But the resources were back up and running on the other node in less than 3 minutes with no administrative action needed.

According to my contacts at MS Exchange2003 is really just a support pack to Exchange2000, atleast in what it changes. The memory management issues with active-active clusters still exists and will not be fixed until Exchange moves away from Jet and moves onto the SQL data storage system. Most of the changes to 2003 are default configuration related, and some ESE Jet improvements that most people will never use. RPC over HTTP is neat, but definitely not a huge I gotta upgrade now kind of thing. Shadow volume copy is probably the best thing in the upgrade, but that isn't going to keep your cluster availability higher.

I want to mention another "solution", that might fit into the
starting question. It might be that I understand "Cluster[ing]"
in a different context.

I'd like to refer to MPI[1] - the message passing interface.
If you, Shugart, have programs available or if you intend to write them yourself,
you can use this interface to let your programs run making use of all available
processors - IF the MPI is installed on each machine.
There is some initial effort, but it pays off. Since years, I am using
large-scale clusters >100 dual-nodes, with all kind of processors
and OSs. (legally! ;) )

Of couse, you also can run a "master"-program on one arbitrary node.

[1] http://www-unix.mcs.anl.gov/mpi/mpich2/

mohaughn.

You have never had problems with your Win2K cluster corrupting Exchange? Microsoft's Exhcnage project manager himself has even apologized for it in public. In fact they no longer recomend an Exchange 2000 Cluster on windows 2000 unless it's very specific in configuration that they trust and that is not Active/Active or even Active/Passive it just does not work and this "sucks ass" in that situation. I guess a file share or load balancing wouldn't be much of a problem or a non-cluster aware virtual server with data stored on the cluster... and I am not familiar with clustering SQL, that might be king.

I admit I was one of the early adopters and have applied a ton of patches to both the OS, motherboard, SCSI firmware, and on and on.... and haven't seen a crash in a few months but now 2003 is out thus my comment. ;) I trust it running but the issues occured while testing failover and in every test the thing failed. AND that is why I bought and configured it.

Configuration issue? What else could it be? :D A faulty MS product implemented on name brand PCs? I blame Dell anyway, and more specifically their POS PERC cards.


Thanks for the Exchange info I too don't think the upgrade is worth it, otherwise I wouldn't be posting. he he, RPC of HTTP is coolness in a bottle. I don't forsee changes to Exchange until the DFS in next generation OS platforms?

So I'll retract my statment or add too it... Clustering sucks azz with Echange.

Quote:

---

You have never had problems with your Win2K cluster corrupting Exchange? Microsoft's Exhcnage project manager himself has even apologized for it in public. In fact they no longer recomend an Exchange 2000 Cluster on windows 2000 unless it's very specific in configuration that they trust and that is not Active/Active or even Active/Passive it just does not work and this "sucks ass" in that situation. I guess a file share or load balancing wouldn't be much of a problem or a non-cluster aware virtual server with data stored on the cluster... and I am not familiar with clustering SQL, that might be king.

---

I highly doubt that anybody at MS has said that Win2k and Exchange2k no longer supports clustering. I work with MS on a daily basis, I actually know several of the critical problem resolution technicians on a personnal basis because of the close interaction we have had with them over the last 3 years. I've not had any problems with MS clustering, specifically disk issues since the clusdisk.sys patches were released almost 3 years ago. Even then we were not losing data, we were just losing disk signatures, which is easily corrected. Active/active clusters have never been supported in Exchange2000 as they have a lot of memory issues when the concurrent connection number goes over 1000.

One of the big problems with clustering is the lack of great documentation for it. I have about 40 pages of technical procedures that I have written that all revolve around using a win2k cluster and exchange. Most of the information came directly from MS, but I've not seen most of the stuff that I have properly documented on the MS technet site. The order in which you start/stop/failover nodes can impact your disk structures. There is a pretty good chance that if you bounce both nodes of a cluster at the same time they will corrupt the disk signatures in the cluster metabase. So yeah, there is a lot of potential to mess things up, but not if you know the product well.

MS does not official support any clustering configuration unless it is very specific in nature, that is nothing new. There are certain firmware revisions, and hardware devices that will not work. If you are not running pretty much the latest and greatest firmware and BIOS revisions on all of your networking components and SAN components, which all have to be officially supported, you will have disk issues. But that is not to say that the product cannot be made stable, or is not recommended. I've never used Dell servers in clusters. Compaq/HP is the preferred vendor for datacenter and clustering and we use their hardware exclusively. Is the product extremely scalable and stable out of the box? Not really, but if you are going to try and run more than 5000 clients per cluster with over 1TB of data and the need to back all of the up in a reasonable amount of time you had better know the product inside and out. Running a default configuration in that instance would not make sense at all. It's not that difficult to tweak the disk and memory configurations in such a manner to make the system stable.

I know the exchange 2000/2003 product manager from being a member of the Exchange 2000 Joint Development partner program. We are still a member of JDP although we did not participate in the early adopter program for 2003 like we did for 2000. If you are a member of the Exchange JPD/dogfood program let me know and I will tell you who I am as I post on that list sometimes.

Quote:

---

MS does not official support any clustering configuration unless it is very specific in nature, that is nothing new.

---

That's all I am saying. I was SOLD on this thing by Microsoft and Dell for it's ability to keep juggin' along with failover. Both Claiming it was tested and "certified" The issue is, it's so touchy it almost never fails over and stays alive. And that is the sole reason I got it. Sure I can be very careful and fail the nodes over properly but what good is that? I grindge at any sort of minor error on the thing and it's Supposed to be fault tollerent. LOL. That is why you buy a cluster outside of distributing work load. I am with you on the lack of documentation. There is some more out now but back 2 years ago I had to feel my way in the dark. So I have to wonder if MS clustering 1. Sucks and 2. If other clustering models are better?

I am not on the list, but I did speak face to face with the product manager over Exchange 2003 and Outlook 2003 at a conference in Denver. It was his own words i was reiterating.

Oh and I do spend a lot of time taking it down (note not fault tolerance) and updating the firmware of a gazillion devices to keep up with it. :) If you don't take it down and update 1 node then chances are very good something will get corrupted so you take the whole thing down update 1 node, power it off, update the other then power the controlling node back up, then the passive and CROSS your fingers. That in my humble opinion is a waste and denys the very reason why clustering is useful outside of load balancing.

We are on the same page just a different book. :)