Cheap cluster for computer chess?

[wgarvin]

Commodity components get cheaper every year, and it seems like a small cluster built specifically for chess engines would be reasonably affordable now. I was curious what it would cost so I priced out some of the components, starting with two assumptions: (1) quad core CPUs, and (2) it would be nice to have motherboards smaller than a full-size ATX board.

All prices are in Canadian dollars.

Lets start with a quad CPU:
$119.99 : AMD ADX635WFK42GI Athlon II X4 635 Quad Core Processor - 2.90GHz, Socket AM3, 2MB Cache, 2000MHz (4000 MT/s), OEM packaging

Those could go into any Socket AM3 motherboard, such as either of these micro-ATX boards:
$64.99 : Gigabyte GA-MA78LM-S2H AMD760G mATX Socket AM3/AM2+, PCI-E, DDR2, SATA2, Video, Sound, GBLAN, HDMI
or
$65.99 : MSI GF615M-P33 Motherboard - NVIDIA GeForce 6150SE & nForce 430, Socket AM3, MicroATX, Audio, Video, PCI Express, 10/100/1000 Ethernet LAN, USB 2.0, RAID

Then you need some RAM for them. Assuming the MSI board, you need DDR3, for example:
2x2GB for $117.99 (after rebate?) OCZ Fatal1ty Dual Channel 4096MB PC10666 DDR3 1333MHz Memory (2 x 2048MB)
or
1x2GB for $69.99 Kingston PC10600 1333MHz 2GB DDR3 Triple Channel Memory
or
1x4GB for $129.97 OCZ OCZ3G1333LV4G Gold 4GB PC10666 DDR3 Memory Upgrade - 1333MHz, Non-ECC, Unbuffered, 1x4096MB

The Gigabyte motherboard would want DDR2, for which prices are even lower.

Okay, now you need a power supply for each node, such as:
320w for $26.99 : Diablotek DA Series 320w MATX Power Supply
or
380w for $35.99 : Diablotek PHD380M 380-Watt MATX Power Supply


Anyway, assume 4 x $120 for CPU, 4 x $66 for the MSI boards, 4 x $118 for 2x2GB of DDR3 RAM and 4 x $27 for power supply. You just bought a 4-node, 16-core compute cluster with 1GB RAM per core, for $1320 plus tax! If you include the cost of a hard drive and some mounting materials and a few short ethernet cables, the total probably comes to around $1500, or 94 dollars per core. For an extra 50 bucks you could use the 1x4GB RAM leaving the option to later double the RAM of all nodes for another 520 bucks.

I didn't give much thought yet to how to mount and cool such a beast. But in the past, others have built "micro beowulf" clusters of this kind and posted photos of them, so that should give some ideas! this one or this one, for example.

[jshriver]

I was hoping to do the exact same thing since it's more cost effective to have a 4 board system with a quad core, than say a really expensive quad socket, with quad cores.

In the end unless you plan on using it daily or lots of testing. I've found it's cheaper to use my home computer for testing, and plan on renting cpu time from a local super computing center. Just recently found out I could get a consumer grade package from them.

Regardless would be a beautiful beast to have
-Josh

[Dann Corbit]

Here is the issue I see with cluster computation:
Where is the break even over conventional SMP?

IOW, if I have a conventional SMP machine with (e.g. 8 cores) what hardware cost for a cluster will give me identical performance. And at what point does clustering become cheaper than SMP (or does it ever?)

What clustering does (IMO) is raised the ceiling if you have either a big pile of cash or a huge pile of hardware at your disposal. So, for instance, Dr. Hyatt's big beowulf cluster makes clustering attractive. Or if you are an oil magnate and would like to build something with one billion NPS net throughput then clustering is the best option since conventional hardware can't get you there. But I guess that for "the rest of us" traditional SMP will probably be a cheaper way to crunch chess.

[jshriver]

I agree, and I dont think clustering will ever surpass SMP due to net latency. Think even with myrinet the latency is still bad compared to SMP.

If I was rich and didn't care about $$ I'd go for a quad socket, with quad cores with dedicated ram channels for each socket.

http://www.ewiz.com/detail.php?p=MB-H8QM32&c=pw
I can dream eh?

Then again I'm not a professional and may stand corrected.
-Josh

[Dann Corbit]

I agree, and I dont think clustering will ever surpass SMP due to net latency. Think even with myrinet the latency is still bad compared to SMP.

If I was rich and didn't care about $$ I'd go for a quad socket, with quad cores with dedicated ram channels for each socket.

http://www.ewiz.com/detail.php?p=MB-H8QM32&c=pw
I can dream eh?

Then again I'm not a professional and may stand corrected.
-Josh

There are also 6/8/12 core models now:
http://www.anandtech.com/show/2978/amd- ... ore-xeon/3

[Gian-Carlo Pascutto]

It seems more cost effective to build it with 6-core Phenom II's now. The CPU is 100 bucks more, but in the end you save on the MB, PSU and RAM.

And more CPUs per node of course helps efficiency.

Magny Cours is interesting if you look at the lowest end CPU, an 8-core 2.1Ghz which only costs about 270 USD. But the mainboard and RAM make it more expensive than the above.

[Gian-Carlo Pascutto]

Here is the issue I see with cluster computation:
Where is the break even over conventional SMP?

IOW, if I have a conventional SMP machine with (e.g. 8 cores) what hardware cost for a cluster will give me identical performance. And at what point does clustering become cheaper than SMP (or does it ever?)

This is entirely dependent on the algorithm. If you are for example playing testgames, scaling is 100% on a cluster.