F@H - Best config for 4-way Xeon HT

[stitchface]

My fellow Poopflinger and I have been discussing the best way to configure some 4-way Xeon servers with hyperthreading to get the most out of them during the short time we'll have them folding. The information available out there is how to configure a hyperthreaded and/or dual machine, but not a 4-way HT server. Please let me know if you've seen any best practices for this situation. Without having a lot of time to test and find the best config, we figured we'd just let the 4 processes (machineID's 1-4) have at it without configuring the affinity for each process.

Thanks,
Stitchface

 

[ProviaFan]

Do you mean a server with 4 HT Xeons (an 8 virtual CPU server) or a server with 2 HT Xeons (a 4 virtual CPU server)? If the former, unfortunately you're out of luck in terms of being able to use all 8 virtual processors, because F@H on Windows only supports up to 4 (I wish the developers would remove this stupid and pointless limitation, but they haven't got a Round Tuit yet).

Take note that I have no personal experience with HT setups and F@H, but from the reports of those who have, it seems that two clients per processor with HT can offer a very small performance over one client per processor. However, some have reported issues with the two clients per HT processor causing other applications to run more slowly (apparently the OS gets confused and ends up giving the two F@H processes more priority than they're supposed to have).

 

[stitchface]

It's an 8 virtual CPU server. Four physical CPUs. I'd heard the same thing about hyperthreading not offering much of a speed bump, but since we plan on leaving HT turned on while it's folding (it won't be doing anything else so other applications won't matter) it still seems like there should be a recommended way of configuring it. Either starting the four folding processes and let them share all the CPUs, or if the affinity should be set so each folding process has a physical CPU all to itself (which is what is recommended for multi CPU machines I think).

For setting up affinity you are given virtual CPU 0-7. I've never found anything that tells you how those CPUs correspond to the actual physical CPUs and without that there is no way to know how to give each folding process it's own physical CPU. In other words, is it CPU 0 and 1, then 2 and 3, etc. that correspond to a physical CPU, or is it CPU 0 and 4, and 1 and 5, etc. On a machine with 4 or less virtual CPUs it doesn't matter of course, just assign each folding process to one of the virtual CPUs.

In my head I calculated that having five 4-way servers folding for 3 weeks is worth over a years worth of processing on a single CPU machine, so a little tweak here or there could be worth some serious beans! Maybe I'll try the Stanford F@H forum and see if anyone over there has heard of recommendations.

 

[ProviaFan]

Originally posted by: stitchface
Maybe I'll try the Stanford F@H forum and see if anyone over there has heard of recommendations.

That's probably the best idea. Of course, if you can run Linux on the servers, then you don't have to deal with the very limiting machine IDs (no Windows Registry to screw things up, either ), and you can easily run 8 clients by having them in 8 separate folders.

 

[Confused]

I believe it's 01 23 45 67 for the 4 processors. You could set up F@H and something else like SoB. If you can get F@H to use 0246 and using the service install of SB, you can set that to use 1357, and you will get the best for both projects


A nice machine to play with though!!!


Confused