Project "The Wall" - dedicated i7 Cluster

[jcool]

Hey folks,

I was waiting on whether FUGGER would dig up some working backups, but apparently that ain't happening any more, so here I go.. repost with some minor updates



The idea here is to create the perfect array of crunchers, within a budget, main focus on high performance (obviously..) while maintaining low running costs (elec...) and also pleasing aesthetics, since in the end this will go on my living room wall (yeah, I know ) which is also the reason I named the project "The Wall".

The nodes will be stacked up on top of each other, for starters I'll go for a 2 wide 2 high config. If those 4 nodes are finished, stage 1 will be complete. Ultimately I am thinking about a stage 2 featuring 9 nodes (3 wide 3 high), but that's somewhere in the more distant future I suppose.

Now, let me show you what I got right now, it's all the parts I need for node #1:



Pretty much the main parts... Next, some close-ups:





Full HW list, though you may have guessed what it is already

- Case: LianLi V350 µATX Aluminum Cube
- CPU: Core i7 920 C0 for now, will be replaced with a W3520 D0 in a week
- CPU Cooler: Scythe Zipang. It features 6 heatpipes and a 140mm (!) fan, definitely the biggest heatsink you can fit in that case. It'll blow the heat directly into the PSU which will then transfer it out the rear, meaning no heat should stay in the case for long. Distance CPU fan <-> PSU fan is like 1-2cm xD
- Mobo: DFI X58 Lanparty JR, a real piece of art. 6 DIMM slots, SLI & Crossfire capability, decent cooling/VRM along with ALL the OC options you can dream of, Diag-LED etc, all slapped on a perfectly designed µATX PCB. Kudos to DFI for making it happen
- VGA: a 55nm Palit/Gainward layout GTX260-216 running on GPUGrid (sadly it's a bad OCer, 650/1450 core/shader is all it'll do) - obviously not the card on the pic
- Ram: Very old OCZ Reaper DDR3-1333 rated for CAS6-6-6 at 1,75V. I had them for like forever, had to swap them out because some Asus 790i didn't like them.. at all Initial tests show they run fine at 1500Mhz CAS7-7-7 with 1,6V
- SSD: Yeah.. no HDDs for a node that state-of-the-art. It's a 32GB Patriot Warp V2 SSD I have no other use for right now (from an RMA, the other SSD from the Raid 0 array died)
- Power: a Seasonic S12-II 430 unit will provide efficient and clean power

The plan is to build 1 node per month, so I can start trying to form a cluster next month. I'm gonna need some help with that, since I have no clue how to do that yet. I am guessing I will need to do some kind of Linux l33t h4xXx0r install, but I'm counting on you guys to help me out when the time comes

---------------------------------------------------------------

Node specs summary:

PPD: 25k WCG at 3,4Ghz --> 30k with W3520
Power draw: 190W headless, 300W with GPUGrid
Space requirements: 262x279x373mm
Noise level: below 22/25 dbA idle/load

[Gamekiller]

At first I thought that the SSD came with the PSU. You can imagine my thoughts...

[123bob]

I know way less than you do about clusters, so I'll ask an obvious noob question. Sorry.

Is the advantage of doing this to create a set of machines that appear as one logical machine to WCG? Is greater crunching efficiency gained somehow? You certainly will rule the machine point producer ranking with it...

Regards,
Bob

[jcool]

I know way less than you do about clusters, so I'll ask an obvious noob question. Sorry.

Is the advantage of doing this to create a set of machines that appear as one logical machine to WCG? Is greater crunching efficiency gained somehow? You certainly will rule the machine point producer ranking with it...

Regards,
Bob

You can't know less than me about clusters, since I already know nothing about them
Except maybe how to hook up the HW...

Why I want to do this? Several reasons.. I've never done it before so I'd like to try, also opening up some taskmgr and seeing 32 cores would sure help make my day
And then there's the coolness factor of running a cluster and also having 1 logical beast in the stats ofc

There won't be any efficiency gains, in fact the goal here must be to avoid any efficiency losses compared to running all the nodes individually.

[shadowwind]

im liking the setup but in my case im looking for a more open box like the antec 300 to run 2 gpu s heat can be a factor in that little case also im thinking of putting in a server cpu like the E 5504 I7 it is 80 watts cpu so that should cut some heat down some i just need some memory suggestions. good luck on your build

[Jaco]

im liking the setup but in my case im looking for a more open box like the antec 300 to run 2 gpu s heat can be a factor in that little case also im thinking of putting in a server cpu like the E 5504 I7 it is 80 watts cpu so that should cut some heat down some i just need some memory suggestions. good luck on your build

That cpu is lacking HT. (and also 4MB L3 cache)
The E5520 has HT and the full 8MB L3 cache.

[shadowwind]

thanks i just added the E5520 to the cart i just need to figure out what memory will work in that little DFI board. im a noob with Intel all my stuff is Amd.

[jcool]

The DFI takes any kind of DDR3, it is a standard desktop motherboard, after all.
I wouldn't go with a E5520 in there by the way. The W3520 Xeon is a lot cheaper, will OC higher, or can OC + undervolt so you can get the TDP way down.

[PoppaGeek]

I have done some cluster work. While I am not sure what you mean by cluster I can tell you the true definition of cluster requires the application to have been coded to run on a cluster. WCG is not. WCG is a grid.

I was a Sr UNIX Systems Admin, Solaris, AIX, Linux, FreeBSD and SCO. I'll help where I can and can do some research for you if you define what exactly you want to accomplish. There are cluster like tools where you can issue commands on one machine that execute on others. But they are not a true cluster just control and management tools for multiple machines.

My plans are to try and setup a group of diskless machines running Linux that are all controlled from one machine all running WCG. But that is still not a true cluster.

Hope my explanation helps.

I love the stacked cubes idea.

[PeterStoba]

I wondered where this had gone Nice idea, would look cool with lots of cubs in the corner of the room

I can't help out with the cluster problems though, sorry

[EvoCarlos]

i found this it may be worth a read
http://www.scl.ameslab.gov/Projects/...ara_intro.html

[jcool]

I have done some cluster work. While I am not sure what you mean by cluster I can tell you the true definition of cluster requires the application to have been coded to run on a cluster. WCG is not. WCG is a grid.

I was a Sr UNIX Systems Admin, Solaris, AIX, Linux, FreeBSD and SCO. I'll help where I can and can do some research for you if you define what exactly you want to accomplish. There are cluster like tools where you can issue commands on one machine that execute on others. But they are not a true cluster just control and management tools for multiple machines.

My plans are to try and setup a group of diskless machines running Linux that are all controlled from one machine all running WCG. But that is still not a true cluster.

Hope my explanation helps.

I love the stacked cubes idea.

Hehe, excuse my noobness with using the word "cluster" then
What's the difference between cluster and grid computing anyway?

What I want to "accomplish" is 4 identical cubes (=nodes) to be seen as one, unified computer (=cluster?). As in, I run one OS on all 4 simultaneously and therefore have 32 threads at my disposal from a single machine.
The machines would be linked via a separate gigabit switch (fibre channel is way too expensive and probably way overkill for this as well).
Now, that OS/Software would then obviously distribute the work to each machine, but the WCG client would think it runs 32 threads on one machine. Much like a HW raidcon that stripes data onto multiple disks, which the apps don't know of course.

Make any sense to you?

[Martijn]

If you've got 2 RJ45 connectors on the boards, you could use both of them. Any decent OS will divide multiple streams of data over the two connections automatically. I guess clustering takes a good deal of bandwidth.

OT (since this is WCG ): heard anything from CSX yet?

[jcool]

No, seems my contact is on vacation or something...
Mini-DFI has only 1 lan port so I guess teaming won't be possible. From what I've read, it's not required anyways if all you do is distribute the work. However if you like render something you'll need a lot of bandwidth.

[Martijn]

No, seems my contact is on vacation or something...
Mini-DFI has only 1 lan port so I guess teaming won't be possible. From what I've read, it's not required anyways if all you do is distribute the work. However if you like render something you'll need a lot of bandwidth.

Maybe toss in some cheap PCIe NICs? Oh well, you could always do that if you need more bandwidth.

[Plan.B]

i found this it may be worth a read
http://www.scl.ameslab.gov/Projects/...ara_intro.html

yes, good link. I've been interested in clusters or "Beowulfs" for awhile but haven't been able to climb the Linux learning curve yet - due to lack of time - at least that's my excuse so far. I've been working with the Kubuntu distro, and have literally pallets of dell optiplex 240's, 260's, and 280's, in various states of functioning or completeness. Keep up the good work on your project. Hopefully I will be able to contribute substantially later.

[xVeinx]

Grid computing is the ability to run multiple computations via network on many different computers. Each WU is it's own instance however. Cluster computing would be closer to running the F@H smp client across all the nodes in your cluster; rather than having instances, you have one instance of the program, with all the cores working together on that single computation (more or less).

You can set up a cluster, but you won't gain any advantage (most likely) simply because you are running an instance on each core, as opposed to using multiple cores for each instance. Regardless, you can still set up a cluster and do that.

[rcofell]

Good luck on your endeavors of making a cluster

Personally, I've been meaning to get a diskless grid running for quite some time; I even had WCG crunch for a month or so on a diskless computer running Linux over NFS two years ago, but school/work pretty much stole all my attention so I haven't been able to get any further. But hopefully things go better this summer/fall, since I'm on co-op at IBM for a full 6-months this time and don't have to necessarily be in a rush to figure out/get my project done.

The way I envision it is I'd run all of my nodes off of a base OS filesystem(read-only via unionFS, including all binaries to be shared), mount any temporary data directories as tmpFS, and have any persistent data(ie. boinc data) go to its own mount over NFS. It wasn't too terribly hard to get everything above in place to the level required for crunching, so the place where I'm looking at putting the most work is writing some Perl scripts(maybe some shell script/AWK as well for good measure :P) to automate adding/removing/configuring nodes, updating the OS filesystem, allowing for specifically tuned kernels, etc. Originally I was thinking about the issue of having 32-bit machines co-exist with 64-bit ones, but nowadays I don't run any machines that're 32-bit only, except my sammy which'll probably get firewall/remote access duty. With that problem gone I'd ideally like it such that I could even plug in a box into the subnet/VLAN and have it automate the configuration to get WCG running under my account without any interaction.

There's also a Linux distribution(Ubuntu based -_-) for diskless crunching that was created by Dotsch on the BOINC forums, which can be found here. I will have to say that I haven't looked into it beyond quickly glancing through the website, since I feel that I started this project a good time before his release and I'd like to keep my ideas original, see where I can take it

[PoppaGeek]

I've seen cluster used incorrectly in the past. What xVeinx wrote pretty much sums it up. Clusters are also called parallel computing which means computations are run in parallel on a cluster of machines. Whereas WCG uses each core to run one instance a cluster uses all cores for each instance. Instance=WU.

Since I do not know for a fact how WCG determines how many cores a system has to determine how many instances to run I am not sure how it will run on a cluster. I think it will determine how many cores are available and then start that many WUs and run across the network on the nodes.

If you want to do this for the learning experience or the cool factor, both legitimate reason IMO, then try it. Knoppix might be the easiest to try it and see if it is what you want to do. I looked at the link rcofell provided and I think I will give it a try. Running off a USB stick sounds interesting and cheap.

I ran Distributed.net on a FreeBSD cluster of 6 Pentium Pro 180mz with a dual Pentium 166 as the proxy/gateway a long time ago. It was fun and I learned alot. Just be prepaired it will take some time to set up and get running correctly.

[SoulsCollective]

I think the term you're looking for here is a farm, not a cluster. Nevertheless, impressive amount of resources you'll have at your disposal, and I'm in awe at the amount of dough you've thrown into this project.

[jcool]

Thanks for all the input guys, I appreciate it

Gonna check out that link rcofell posted, put I'll probably gonna have to ice this cluster (or whatever the word is ) business for a while, due to lack of time and also a 2nd node to experiment with.
Seeing as I spent money that I didn't really have on a Dual Gainstown build, node #2 will probably be a while.

[jspace]

I have done some cluster work. While I am not sure what you mean by cluster I can tell you the true definition of cluster requires the application to have been coded to run on a cluster. WCG is not. WCG is a grid.

I was a Sr UNIX Systems Admin, Solaris, AIX, Linux, FreeBSD and SCO. I'll help where I can and can do some research for you if you define what exactly you want to accomplish. There are cluster like tools where you can issue commands on one machine that execute on others. But they are not a true cluster just control and management tools for multiple machines.

My plans are to try and setup a group of diskless machines running Linux that are all controlled from one machine all running WCG. But that is still not a true cluster.

Hope my explanation helps.

I love the stacked cubes idea.

Some of the research I have done on this points to the need for SSI (Single System Image) to work. Do you know anything about those kinds of setups?

[PoppaGeek]

Some of the research I have done on this points to the need for SSI (Single System Image) to work. Do you know anything about those kinds of setups?

Running a common root filesystem with each instance having own home file system?

[jcool]

Update, changed the 920 C0 for a w3520, batch 3848A352:



Looks very good a first sight, however this thing is the most evil low vid CPU I have seen since my old B3 Xeon 3210.
At that insanely low Vcore, it takes 260W BOINC loaded with no GPU, 380W with the GTX 260 running GPUgrid. Temps are insane too, like 85C on all cores, 90 while running prime. Still the Scythe Zipang, contact isn't bad, I checked it - however it seems stable so I'll let it run for now.

[haschioz]

well, jcool thats true a low vid!
just a pity you cant fit a true in there, but nevertheless i like your idea and the style