F@H running on a cluster?

[Shambler]

How would this work? Single client on mother node sending work to all children node... Is there a Cluster Client?

Would a couple old machines clustered yield better results than a clocked Quad? (Excluding power per watt, and folding per watt ratios)?

Has anyone tossed up a couple Quads into a cluster for folding?

I don't have the budget, but I damn well have the imagination! I wonder if Stanford has individual machines folding, or has 'em clustered...


Your thoughts

 

[alan2308]

No cluster client. However, I heard somebody at Stanford would be interested in working on one if someone had the hardware.

 

[Sunin]

My guess would be that you would lose efficiency in overhead versus running your quads with just two smp's. For the same reason that 1 smp does not full saturate a Quad. 2 SMP's are required. Thus a Cluster would need multiples and I think if say you clustered 4 P4's together you would be able to run a single SMP only. But if you clustered 4 quads you would need 8 smp's and the risk that one smp craps out and takes the other 7 with it honestly would be too much of a risk. I've had issues even using the power shutdown button with the SMP client deleting a work unit! This issue I would think would be exponentially worse with a cluster.

 

[alan2308]

My guess would be that you would lose efficiency in overhead versus running your quads with just two smp's. For the same reason that 1 smp does not full saturate a Quad. 2 SMP's are required. Thus a Cluster would need multiples and I think if say you clustered 4 P4's together you would be able to run a single SMP only. But if you clustered 4 quads you would need 8 smp's and the risk that one smp craps out and takes the other 7 with it honestly would be too much of a risk. I've had issues even using the power shutdown button with the SMP client deleting a work unit! This issue I would think would be exponentially worse with a cluster.

You'd also need some pretty serious networking gear to make it worthwhile. And for that money you could pick up a few more quads.

Also, that offer from Stanford has been on the table for a couple years now. And nobody has taken them up on it yet.

 

[-Sn1PeR-]

Does 1 SMP client really not fully saturate a quad core? Mine appears to be...

 

[owkia]

Does 1 SMP client really not fully saturate a quad core? Mine appears to be...

You'll get more PPD if you run two so no, one client doesn't fully saturate a quad.

 

[Tigerbiten]

Does 1 SMP client really not fully saturate a quad core? Mine appears to be...

The gain from going from 1 client to 2 clients on a Quad box is around 20-30% depending on the protien.

Take p2605 for example.
It runs at 250 PpDpGhz on a 4 core and at 330 PpDpGhz on 2.
On a 3 Ghz, 4 core box, one copy gives ..........
250 PpDpG x 4 Cores x 1 Copy x 3 Ghz = ~3,000 PpD.
On a 3 Ghz, 4 core box, two copies gives ............
330 PpDpG x 2 Cores x 2 Copies x 3 Ghz = ~ 4,000 PpD.

Thats the reason I've got 12 CPU's but are running 24 clients.
But the only thing about running twin clients on a CPU is you get them both back well before the preferred deadline so the don't get re-issued.

Luck ...........

 

[Kendrak]

The gain from going from 1 client to 2 clients on a Quad box is around 20-30% depending on the protien.

Take p2605 for example.
It runs at 250 PpDpGhz on a 4 core and at 330 PpDpGhz on 2.
On a 3 Ghz, 4 core box, one copy gives ..........
250 PpDpG x 4 Cores x 1 Copy x 3 Ghz = ~3,000 PpD.
On a 3 Ghz, 4 core box, two copies gives ............
330 PpDpG x 2 Cores x 2 Copies x 3 Ghz = ~ 4,000 PpD.

Thats the reason I've got 12 CPU's but are running 24 clients.
But the only thing about running twin clients on a CPU is you get them both back well before the preferred deadline so the don't get re-issued.

Luck ...........

Do you have some massive spreadsheet that you keep all this info in? Or do you just crunch it on the fly?

 

[capreppy]

Do you have some massive spreadsheet that you keep all this info in? Or do you just crunch it on the fly?

I think those are pretty standardized numbers. I'm getting 330 PpDpGhz on my Q6600 @ 3.42. I am running dual Linux VMs and add a standard console client to take up any slack. I get pretty close to 4550 PpD.

 

[-Sn1PeR-]

Do I need to install two different copies of F@H to do this, or just boot up the smp exe twice?

 

[alan2308]

Do I need to install two different copies of F@H to do this, or just boot up the smp exe twice?

Two different copies from two seperate folders. You also need to change the machine ID to 2 for the second.

 

[capreppy]

You will need to install the SMP client to TWO different folders. I use F@H1 & F@H2. Not very original, but it works for me.

 

[APOLLO]

You'd also need some pretty serious networking gear to make it worthwhile.

Such as 10GbE or Infiniband, etc. Mucho dinero.

And for that money you could pick up a few more quads.

Yes, precisely why there's no way I see someone actually building a cluster strictly for F@H. The money saved in the networking infrastructure alone could increase your production an order of magnitude if allocated to purchasing additional real hardware (D).

Also, that offer from Stanford has been on the table for a couple years now. And nobody has taken them up on it yet.

Probably because their existing clusters are doing work they view as much more important to them, or relevant to their research.