Folding time. Finally?

[CMetaphor]

Hello all. Im thinking its about time to get my server folding away. Was re-reading the smp/bigadv setup guide when I read this little tidbit:
"SMP requires a minimum of .5 GB per core"

Um. Uh oh... My server doesn't have .5Gb per core currently, does that mean it cannot possibly fold properly?

Also: En lieu of partition my Server2008R2 OS drive to make room for a linux one JUST for folding, is there any F@H live CDs, or better yet persistent USB key linux distros, that will boot up up straight into folding by itself? Yes. I'm that lazy... hah.

[great_big_abyss]

I don't think they mean .5GB / CPU core, I think they mean .5GB per FAHCORE, or 'instance'. For example, you may have a 64 threads, and 16GB of RAM, therefore .25GB of RAM per CPU Core, but only one FAHCORE running across all 64 threads, therefore you are okay.

Somebody correct me if I'm wrong.

[SugarJ]

Use a VM for your linux folding.

[CMetaphor]

Originally Posted by SugarJ

Use a VM for your linux folding.

I thought that was no longer a requirement? So much for F@H then... cant even make it properly scare above 16 threads without the use of VM

[SugarJ]

It's not a requirement at all. Just the easiest way to get linux folding on your windows box without constant rebooting.

Have a look at this page, there's some linux distros that you should be able to put on a USB key. Third Party Contributions - FaHWiki

[Dead Things]

Originally Posted by great_big_abyss

I don't think they mean .5GB / CPU core

That's actually exactly what they mean. But it's a rule of thumb that only really applies to desktop hardware. I don't have a 64-thread machine, but I'll check my 48- and 80-thread boxes tonight to find out how much address space each needs by the end of a work unit. IIRC, 6900-series work units needed between 4-6GB by the end of the unit, so that is probably roughly doubled with the current 8100-series work units. But as I said, I'll confirm that for you tonight.

In your case, virtual machines are not an option I'm afraid. If performance on bare metal Linux is 100%, then...

• Performance on a Linux VM with one socket is about 90%.
• Performance on a Linux VM with two sockets is about 70% (i.e., better than running separate instances, but only marginally better than running regular SMP on Windows).
• Performance on a Linux VM with four sockets is about 30% (i.e., significantly worse than running separate instances - in fact, only marginally better than running a single instance on one socket).

So your choices are:

• Boot into Linux specifically to fold bigadv units. The pro is higher production while running. The con is that due to the high importance of quick turnaround with these units, you basically have to turn the machine into a full-time folder to make it worthwhile. Or...
• Fold regular SMP units on Windows. The con is lower production while running. The pro is you can use your machine for whatever you want whenever you want, and since you aren't as much of a slave to the quick return bonus with regular SMP, your overall production will likely be better in the long run.

So unless you are inclined to turn the box into a full-time folding machine on Linux, then I think you can guess what I'd recommend...

Oh, and by the way...

[Dead Things]

Originally Posted by CMetaphor

cant even make it properly scare above 16 threads without the use of VM

Actually fah scales extremely efficiently - even on Windows. Up to 64 threads, that is. Now 10 years after 2003 x64 was released, 64 is still the max thread count per processor group on Windows. Honestly, if not for legacy applications, I can't think of a single reason for maintaining a Windows server. But that's another story...

In terms of efficiency, scaling is in the 95%+ range up to about 256 logical processors. Current fahcores struggle with processor counts over 256, but that is only a theoretical problem at this point. The Xeon Phi at 240 threads is as close as current single-node hardware comes to encountering fah's efficiency ceiling.

The main reason why Linux is preferred over Windows is because of the availability of bigadv units. Otherwise, regular SMP units are almost as fast on Windows as they are on Linux (up to 64 threads).

[Tim_H]

i'll second what DeadThings says. folding is folding, If you don't have a strong need for maximum points then SMP under windows still does good, if you aren't going to switch to linux then windows smp is your best bet.

I can't comment on scaling I haven't tried any testing myself, but from the reading I've done it all supports what has been said already.

in regards to memory, the linux bigadv units take about 3-5GB (48 core bigadv with 32gb ram) while the linux smp units take about 550MB (total system memory taken from a dual lga771 system with 8gb ram)

[Dead Things]

Originally Posted by Tim_H

in regards to memory, the linux bigadv units take about 3-5GB (48 core bigadv with 32gb ram)

Aye - Tim_H is right! Much to my surprise the newer, bigger bigadv projects consume less address space than the older, smaller units. On my 80-threader at 74% into a P8101, fah is using 3.1GB; and one 48-threader at 11% into a P8102 (yay), fah is using 2.4GB; and on the other 48-threader at 17% into a P8101, fah is using 2.4GB.