Windows SMP performance (split from the notfred thread)

Come join the... uh... er... fold.

Moderators: just brew it!, farmpuma

Windows SMP performance (split from the notfred thread)

Postposted on Tue Feb 15, 2011 9:22 pm

There's no value in 4 instances of the SMP client, since the faster the WU is returned the greater value to Stanford, and of course you get more points for it. You (or emkubed as the case would appear to be) can obviously do what I was doing for a number of years, which is running Windows XP (which in my experience is still the best for GPU folding) on a cheap HD and run notfred's SMP client / Linux distro on a VM running under Windows. That's still probably pretty efficient, and except for the forever-taking-to-fix problem of the Linux SMP client not allowing -bigadv for the time being, it's probably the way to go for most machines unless they've got >= 8 cores or hyperthreads or whatever. Only Windows does the GPU client, and in the one machine I tried to get it running under wine, it was a bigger PITA than it was worth, since only particular GPU's allowed you to modify their clocks. Windows is simpler because of the utilities available and the fact that it's over 80% of everything out there.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Re: New Notfred diskless version out

Postposted on Tue Feb 15, 2011 10:27 pm

Ragnar Dan wrote:There's no value in 4 instances of the SMP client, since the faster the WU is returned the greater value to Stanford, and of course you get more points for it. You (or emkubed as the case would appear to be) can obviously do what I was doing for a number of years, which is running Windows XP (which in my experience is still the best for GPU folding) on a cheap HD and run notfred's SMP client / Linux distro on a VM running under Windows. That's still probably pretty efficient, and except for the forever-taking-to-fix problem of the Linux SMP client not allowing -bigadv for the time being, it's probably the way to go for most machines unless they've got >= 8 cores or hyperthreads or whatever. Only Windows does the GPU client, and in the one machine I tried to get it running under wine, it was a bigger PITA than it was worth, since only particular GPU's allowed you to modify their clocks. Windows is simpler because of the utilities available and the fact that it's over 80% of everything out there.

If you can run XP and be a VM host, you can just run the latest beta WinSMP, the A3's are pretty tasty and no more mpich/deino crap. Of course then GPU client will be available too, but you'll need a disk.
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24378
Joined: Mon May 24, 2004 2:19 am

Re: New Notfred diskless version out

Postposted on Wed Feb 16, 2011 1:43 pm

I mentioned the Linux under VM thing because I assume it's still more efficient than the latest Windows one, and the only reason to run the Windows client is for its unique -bigadv capacity, as far as I'm concerned. Of course, I haven't done any experiments, and though I think someone has an SMP test utility somewhere (I forget where I saw it mentioned now, maybe someone pointed to it here, and it may have been on HardOCP or on Foldingforum.org), I still doubt Windows is as efficient at SMP as Linux is.

As for the rest, yes, like I said.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Re: New Notfred diskless version out

Postposted on Wed Feb 16, 2011 3:27 pm

Ragnar Dan wrote:I mentioned the Linux under VM thing because I assume it's still more efficient than the latest Windows one, and the only reason to run the Windows client is for its unique -bigadv capacity, as far as I'm concerned. Of course, I haven't done any experiments, and though I think someone has an SMP test utility somewhere (I forget where I saw it mentioned now, maybe someone pointed to it here, and it may have been on HardOCP or on Foldingforum.org), I still doubt Windows is as efficient at SMP as Linux is.

AFAIK the inefficiencies were mainly due to mpich/deino in the A1/A2 cores where they have to go multiprocess. Now for A3 everything is multithreaded Linux and Windows should be the same, with Windows having the advantage of -bigadv. I have no reference data on regular WUs with WinSMP vs LinuxSMP someone else can chime in, but I expect at least par or more performance from WinSMP now (not to mention I can run it now as a service in the background again, like the old single core client).
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24378
Joined: Mon May 24, 2004 2:19 am

Re: New Notfred diskless version out

Postposted on Wed Feb 16, 2011 4:40 pm

Flying Fox wrote:AFAIK the inefficiencies were mainly due to mpich/deino in the A1/A2 cores where they have to go multiprocess. Now for A3 everything is multithreaded Linux and Windows should be the same, with Windows having the advantage of -bigadv. I have no reference data on regular WUs with WinSMP vs LinuxSMP someone else can chime in, but I expect at least par or more performance from WinSMP now (not to mention I can run it now as a service in the background again, like the old single core client).

I probably should have done a semi-clean install (new directory, copy over queue.dat, client.cfg, etc. files, work directory) when I installed it, because I still have those old files lying around there.

But still, it's Windows, and you know how bad Windows is at a lot of things. It doesn't have as good priority control, and of course there's all the crapware and wasted cycles on things people don't notice running, esp. since Windows all too often doesn't even show them in taskman. Not that the crapware wouldn't still be there with a VM running Linux, but things still ran very well for me when I was running my Opteron 165 (dual core bought in late 2005) at ~2.5 GHz with VMware and a full UI-laden Ubuntu distro.

If we were talking about an 8 thread machine, there'd be no argument. But for now, unless someone can find an ISO with all the necessary goodies to run a GPU client under Wine, which should be possible if the user has enough understanding to do some manual setting up, and probably downloads too, it seems like the best is the Windows + Linux under VM if there are < 8 threads available. One bit of evidence I have for that is the wild variation in frame times on my machine even when its otherwise quiescent. I know that can't be blamed on Windows only, but compared to my Q8400 running notfred's ISO, it's all over the place, varying by more than 10%.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Re: New Notfred diskless version out

Postposted on Thu Feb 17, 2011 12:25 pm

Ragnar Dan wrote:But still, it's Windows, and you know how bad Windows is at a lot of things. It doesn't have as good priority control, and of course there's all the crapware and wasted cycles on things people don't notice running, esp. since Windows all too often doesn't even show them in taskman. Not that the crapware wouldn't still be there with a VM running Linux, but things still ran very well for me when I was running my Opteron 165 (dual core bought in late 2005) at ~2.5 GHz with VMware and a full UI-laden Ubuntu distro.

If we were talking about an 8 thread machine, there'd be no argument. But for now, unless someone can find an ISO with all the necessary goodies to run a GPU client under Wine, which should be possible if the user has enough understanding to do some manual setting up, and probably downloads too, it seems like the best is the Windows + Linux under VM if there are < 8 threads available. One bit of evidence I have for that is the wild variation in frame times on my machine even when its otherwise quiescent. I know that can't be blamed on Windows only, but compared to my Q8400 running notfred's ISO, it's all over the place, varying by more than 10%.

That really was the old client. Ever since I used the latest WinSMP beta most of these problems went away. The most important aspect for me is the usability. I have my brother's machine borged from the moment I put it in production. While I could have configure VM-autostart and all that, I did not. So everytime he did a Windows Update or a crashed happened, I lost the VM and I had to constantly babysit and launch it from remote again (that is why I installed VMware server too, to make the VM less obvious). A week ago I finally switched to WinSMP, no more deino/mpich crap with higher mem requirements, stable points, and best of all I can once again run it as a service. This means that it auto-starts in the background nicely after a crash/reboot, and I don't have to care too much about it anymore. This new WinSMP 6.3 client is as close to the good old single core client IMO in terms of stable performance, resilency, and "set it and forget it"-ness.

That said, let me convert another previous VM-running-in-background node to WinSMP, that one is my WHS and runs pretty stable. I can get some better comparison compared to my brother's machine. It needs another day to complete the current WU so will report back.
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24378
Joined: Mon May 24, 2004 2:19 am

Re: New Notfred diskless version out

Postposted on Thu Feb 17, 2011 3:33 pm

If I could get the Windows SMP client to run as a service on Windows 7 (or I assume any 6.x version), I'd probably be a lot less bothered. But as it is, I have to run it all manually for the time being, and Stanford makes it sound as though that won't be changing. Which sucks.

Speaking of Stanford, did I hear about this before, or not? There's a new GPU client released in January, which appears to be the release version of the beta I'm still running. Anyway, it's the same file size but it isn't the same, so they appear to have tweaked it at least a bit. I learned of that on Vijay Pande's blog: http://folding.typepad.com/.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Re: New Notfred diskless version out

Postposted on Thu Feb 17, 2011 7:14 pm

Ragnar Dan wrote:If I could get the Windows SMP client to run as a service on Windows 7 (or I assume any 6.x version), I'd probably be a lot less bothered. But as it is, I have to run it all manually for the time being, and Stanford makes it sound as though that won't be changing. Which sucks.

Your information is quite out of date. Please read the new 6.3 info in the install guide. My brother is running Windows 7 right now, with 2 cores out of his 965BE running WinSMP. I just did not want him to see maxed out CPU and may be reduce some heat output. Really, WinSMP has finally come back in decent shape. It has been close to a couple of weeks. I can report back later for some longer-term explanations, but right now that set up has folded close to a dozen WUs with no issues so far.
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24378
Joined: Mon May 24, 2004 2:19 am

Re: New Notfred diskless version out

Postposted on Thu Feb 17, 2011 10:29 pm

Flying Fox wrote:
Ragnar Dan wrote:If I could get the Windows SMP client to run as a service on Windows 7 (or I assume any 6.x version), I'd probably be a lot less bothered. But as it is, I have to run it all manually for the time being, and Stanford makes it sound as though that won't be changing. Which sucks.

Your information is quite out of date. Please read the new 6.3 info in the install guide. My brother is running Windows 7 right now, with 2 cores out of his 965BE running WinSMP. I just did not want him to see maxed out CPU and may be reduce some heat output. Really, WinSMP has finally come back in decent shape. It has been close to a couple of weeks. I can report back later for some longer-term explanations, but right now that set up has folded close to a dozen WUs with no issues so far.

Hm. I'll have to test the new client if I don't already have it, or at least try installing it as a service, but I recall reading on foldingforum.org not so long ago (or so it seems) how they thought making a service which could interact with the desktop was impossible, and after reading that I pretty much gave up on them. Apparently someone with more patience got through to them.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

Re: New Notfred diskless version out

Postposted on Fri Feb 18, 2011 12:51 am

Ragnar Dan wrote:
Flying Fox wrote:
Ragnar Dan wrote:If I could get the Windows SMP client to run as a service on Windows 7 (or I assume any 6.x version), I'd probably be a lot less bothered. But as it is, I have to run it all manually for the time being, and Stanford makes it sound as though that won't be changing. Which sucks.

Your information is quite out of date. Please read the new 6.3 info in the install guide. My brother is running Windows 7 right now, with 2 cores out of his 965BE running WinSMP. I just did not want him to see maxed out CPU and may be reduce some heat output. Really, WinSMP has finally come back in decent shape. It has been close to a couple of weeks. I can report back later for some longer-term explanations, but right now that set up has folded close to a dozen WUs with no issues so far.

Hm. I'll have to test the new client if I don't already have it, or at least try installing it as a service, but I recall reading on foldingforum.org not so long ago (or so it seems) how they thought making a service which could interact with the desktop was impossible, and after reading that I pretty much gave up on them. Apparently someone with more patience got through to them.

The old service does not interact with the desktop too, what the heck were those posters smoking? Yes, under the post-Vista era, a Windows system service interacting directly with the desktop is not possible without an intermediary process. However, all the folding client service needs is just to write the log files like before, nothing the current beta cannot handle. The updated instructions are pretty clear about that you should run the fah executable under a writable folder (%ProgramFiles(x86)% is a bad idea) and that when you run -config/configonly you need to be UAC-elevated in order to register the service. I followed the instructions and it worked the first time. Try it, you would be surprised. ;)
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24378
Joined: Mon May 24, 2004 2:19 am

New Notfred diskless version out

Postposted on Fri Feb 18, 2011 2:08 pm

Yeah I know, and what's all the more odd is I think it was Pande himself who made the claim. If not him it was a forum mod or even possibly someone who was putatively working on the software. Whatever it was, I remember reading it and thinking something like, "Well, these people won't be making it work properly any time soon. I may as well quit waiting for them." As usual, that's probably around the time it started being (somewhat) straightened out.

I played with making it a service last night after replying, and it... sort of worked. I have to keep reminding myself not to offend the sensitive prissy software and data of this project by looking sideways at it. :evil: Instead of continuiing on the current P6900 WU at 30%, it restarted from 0% and wiped out my progress, all at startup before I could login and stop it (if that would even have been possible). I should've saved the entire directory first instead of trusting it. I'm wondering if file permissions / ACL's during startup weren't reported right or something odd like that, and perhaps a half minute pause before allowing the client to start would make it work properly every time. Even worse, it is running more slowly than the manually started one did. At first, because I also tried to make the GPU client run as a service and it was failing to launch, I thought that was slowing it down, but after finally disabling that it still runs slower, so I may decide to give up the whole idea if things don't improve. But before that, I'm probably going to have to do a reinstall of the entire OS because of the problems the motherboard swap and Windows Security Essentials' "upgrade" have caused, and I'm starting to lose patience with all the screwing around I'm having to do to get it straightened out.

Just to be clear about this, I've been running Win 7 since July, and the service stuff did not work when I first tried it, nor when I tried it again months later with an updated version. So it's not me that was the problem, at least until the recent change that I was unaware of. Still, though, it looks to me like my GTX 260 on Win 7 Pro is always slower than it was on XP Pro. For example, on 353 point WU's, it pretty much always ran 36 seconds a frame after I OC'ed it. On 7 Pro, it peaks at 37 seconds, but generally is at 38 seconds, and now and then 39 seconds. That's rather abysmal if you ask me. But then I haven't updated the drivers in a while and I can never tell which one people think is best since people alwys seem to have wildly varying results.

One more thing: this conversation of ours should probably be split off from notfred's thread.
Ragnar Dan
Gerbil Elder
Silver subscriber
 
 
Posts: 5355
Joined: Sun Jan 20, 2002 7:00 pm

WinSMP

Postposted on Fri Feb 18, 2011 2:19 pm

Ragnar Dan wrote:One more thing: this conversation of ours should probably be split off from notfred's thread.

Yeah, I was thinking the same thing. Can't deal with it right now (I need to run out for a bit), but I will try to deal with this later.
(this space intentionally left blank)
just brew it!
Administrator
Gold subscriber
 
 
Posts: 37672
Joined: Tue Aug 20, 2002 10:51 pm
Location: Somewhere, having a beer

Re: New Notfred diskless version out

Postposted on Fri Feb 18, 2011 8:09 pm

Since I was switching from LinuxSMP in a VM to WinSMP running as service, I -oneunit the LinuxSMP and then did a clean cut to WinSMP. I never expected them to be cross-compatible. It seemed to work for me. I have just migrated to another WinSMP on my WHS computer. That one had a lot of LinuxSMP WUs folded so I would be able to compare points after a while. Will report back.

@JBI: I continue to post until we eventually split this off. This can be the start of a new thread about how well (or not) the new WinSMP is doing.
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24378
Joined: Mon May 24, 2004 2:19 am

Re: Windows SMP performance (split from the notfred thread)

Postposted on Sat Mar 19, 2011 5:57 pm

Have folded a number of WUs now. LinuxSMP under VMware Server v3 vs WinSMP on the Windows host. WinSMP on average seems to be doing 20-30% better ppd wise (based on HFM.net). The only exceptions for me are projects 6701 and 6702 where there was no difference. My brother also noticed higher CPU fan noise under WinSMP on host compared to LinuxSMP "hidden" in a VM, suggesting that it may indeed be using more CPU cycles and generating more heat.

I think I can finally recommend WinSMP now. drfish and leor, are you coming back? ;)
Image
The Model M is not for the faint of heart. You either like them or hate them.

Gerbils unite! Fold for UnitedGerbilNation, team 2630.
Flying Fox
Gerbil God
 
Posts: 24378
Joined: Mon May 24, 2004 2:19 am


Return to TR Distributed Computing Effort

Who is online

Users browsing this forum: No registered users and 2 guests