Linux, Q6600 - one client? two clients?

bassman

[H]ard|Gawd
Joined
Jun 30, 2005
Messages
1,393
I fired up one smp instance on my quad and noticed it only used 50% of each CPU. Does it make sense to run two instances? Any tricks/tips? Any drawbacks?
 
Excellent question. Alan, Xilikon and a few others could give you an excellent answer. Too bad I have no Linux experience to tell you myself. :(

 
I fired up one smp instance on my quad and noticed it only used 50% of each CPU. Does it make sense to run two instances? Any tricks/tips? Any drawbacks?

Did you use the -SMP flag. Without that on my dual core it doesn't max things out (Ubuntu 8.04).
 
While I have experience with Linux folding, I never tried running 2 SMP together because I'm running VMWare so each instance get 1 SMP client and each box get 2 VM instances to maximize the production.

I dunno if there is a issue with corruption when you shutdown one of the two SMP instances so I'll let someone else like SmokeRngs comment on it.

 
Still waiting for my first quad, so I'm not speaking from experience, instead from what I've read here and at other teams forums.

From what I've seen so far, 2 SMP clients is the way to go on Linux as well. However, this is assuming that you have enough memory to run two, and with bigger and bigger units, thats not a given anymore. Also, if this box is your daily driver, two clients will be more noticeable.

And the corruption issue when you shutdown one of the clients doesn't seem to be an issue on linux.
 
Did you use the -SMP flag. Without that on my dual core it doesn't max things out (Ubuntu 8.04).

Yup. I get roughy 50% usage on all 4 cores according to top.

For the hell of it I fired up another SMP instance. Usage was around 85% on every core, so I shut it down to do some research and thinking.

Now, after a couple hours, I noticed the single instance is using 50% of three cores and 90% of the fourth. Seems odd to me.

Xilikon said:
While I have experience with Linux folding, I never tried running 2 SMP together because I'm running VMWare so each instance get 1 SMP client and each box get 2 VM instances to maximize the production.

I'll give that a shot. What are you using to host the VM's?

 
Wow :eek:, now I'm really worried (not really), using the -smp flag in Ubuntu v8.04 I get around a 50% to sometimes 90% + usage on both my Q6600's (each core alternately). When I use the windose WinSMP client (it's been a while since I've used the WinSMP client, all the way since yesterday) I get 100% cpu core usage all the freakin' time (all four cores at the same time, all the time) Does anyone else have this strange occurrence and should I give a try to running 2x instances of LinSMP? (straight, not VM'ed, using two cores for each client ) :confused:

Please help me if my methods are fubared or I'm fubarring my Q6600's, I've been folding on one Q6600 like this for about a month. (single client, LinSMP) :eek::(

Edit: Since the weather warmed up around here and made a believer out of me concerning the "global warming" thing I went thru my spare parts closet, put together a WC'ing loop for my air cooled Q6600 (MCP350 stock top, MCR220, bowed Apogee GT and a pretty ragged Micro rez) and I'm leak checking it in the shower as I do my "one finger" typing method posting this post (I'm also "typing challenged" :p)

I was gonna' use the WC'ing stuff for a GPU if I ever get one.

 
The core used in the linux units that I've seen, FahCore_a1.exe scales TERRIBLY to quads, which is why a lot of us have been running dual virtual machines in Windows to maximize usage. If you found one in Windows running 100% through 4 cores, it may be using the FahCore_a2.exe I've heard rumors of coming out, though I'm not too sure what the schedule is there. Xilikon may have better answers to that end.

 
AFAIK, the A2 core has not been released for mass consumption yet. The Windows client will use all for cores on a quad, but not efficiently. Task manager displays all four CPU graphs at maximum, though, FWIW.
 
Well, I think I can give a bit more info on this deal. Two of my quad boxen that mostly just fold running Linux I have two Linux VMs setup on them for folding. This gives me the best PPD overall because it usually gets the 2605 work units which kick ass for points. My main machine is a different story. At one time I ran dual VMs on it which worked great. However, after updating the nVidia drivers at one point things went to hell after that.

Because of the dual VMs and their reluctance to give up CPU cycles I was lagging like crazy in games. I then switched to dual clients running natively. For gaming and everything else this has worked out great. The only downside is that both clients get "chewier" work units instead of the really fast 2605s and 2653s (mostly just 2605s). My PPD on the machine went from around 5k PPD to around 4K PPD at most.

One thing to remember is that Linux does not have the corruption problems when shutting down the clients like Window does when running two native SMP clients. You can close one of them and the other will keep going with no trouble. I've had several power outages and other things happen and I can count on one hand the number of work units I've lost because of that. The Linux SMP clients are a hell of a lot more stable than the Windows SMP clients.

If you have the machine overclocked to at least 3.2 I would suggest running 2 SMP clients natively for ease of use. Anything slower than that and it might not make deadlines on some of the work units which is bad. I've almost had this happen with my machine at 3.6. As stated previously, some work units are horrible about scaling up to 4 cores which wastes a LOT of CPU cycles. Running dual clients when crunching these work units doesn't help out the PPD a whole lot, it is noticeable. Even with dual clients crunching on the chewy work units, my cores usually don't go much higher than 80% utilization on each core. I'm afraid there isn't much which can be done about that at this point. Most likely we'll just have to wait for the A2 core to be released to an open beta or something which should do much better with scaling with more than 2 cores.

If this doesn't make sense, well I'm a rambling man at times. Just say something and I'll attempt to use drunken English which would probably be easier to understand.

 
I've been running a single SMP client in XP on my Q6600. The sick part is that my E8400 takes the same amount of time to crunch a 2665 as the quad.

Since I've got a second q6600 (arrived yesterday, rebuilding this weekend) I'm looking into running dual VM's... Any good guides for setting that up? What do people use for affinity management in XP? If there are good guides out there feel free to point me to them.

 
Exactly what I was looking for. Thanks!

How does the affinity stuff work? Just let both VM's run and let 'doze sort it out?

 
I always put my VMs on 0+1 and 2+3....windows would probably sort it out more or less alright, but it will be way more efficient (and in the end higher ppd) to just set it yourself.

 
If you run 2 VMWare instances, just open the task manager and set the affinity of each VMX instance to core 0+1 and 2+3.

 
If you have the machine overclocked to at least 3.2 I would suggest running 2 SMP clients natively for ease of use. Anything slower than that and it might not make deadlines on some of the work units which is bad. I've almost had this happen with my machine at 3.6. As stated previously, some work units are horrible about scaling up to 4 cores which wastes a LOT of CPU cycles. Running dual clients when crunching these work units doesn't help out the PPD a whole lot, it is noticeable. Even with dual clients crunching on the chewy work units, my cores usually don't go much higher than 80% utilization on each core. I'm afraid there isn't much which can be done about that at this point. Most likely we'll just have to wait for the A2 core to be released to an open beta or something which should do much better with scaling with more than 2 cores.

If this doesn't make sense, well I'm a rambling man at times. Just say something and I'll attempt to use drunken English which would probably be easier to understand.

Made sense to me. Thanks for the info.
 
Back
Top