As the Work Unit Crunches

[Flying Fox]

The Xeon 3210 is still a collection of parts waiting to be built. However, ages ago, I did briefly run the CPU in my son's mobo and IIRC it ran about 2,000 PPD on the fastest a1 WUs.

My experience seems to indicate that the A3's should give a marginal improvement to the faster A1/A2 WUs, so it remains to be seen. Were that result obtained from WinSMP as well? Using the same FahMon?

The numbers listed above are from my C2D E4400 (65 nm) using the online calculator link that Maph posted - http://www.linuxforge.net/bonuscalc2.php. I'm running WinSMP 6.29 on XP SP3, although I have not yet tried the drop-in binary update.

Now that makes a bit of sense. Dual core at 2.0GHz with 1200ppd vs my 2.8GHz (plus 45nm improvements) at ~1600ppd. I'm sure the quad core should do better, especially in the bonus department since the job should be finished earlier...

The only other things running are Asus PC Probe II, FahMon 2.3.4, and windows explorer.

Not too sure if PC Probe is a CPU hogger, I would turn it off. FahMon 2.3.4? Not the latest (2.3.99.1)? I think the numbers are different between versions too.

Does HFM.net give you better numbers than the linked online calculator?

I think so, the formula may be more up-to-date (Harlam keeps updating), plus I have multiple finished WUs to average out so I have some confidence in the numbers.

I am sensing that your measurements are not uniform, making it more difficult to compare numbers. May be we should all just settle on HFM.net and go from there?

[Ragnar Dan]

http://techreport.com/discussions.x/18962

According to the above link, nVidia 400-series GPU's now have CUDA 3.1-enabled beta drivers. Does this mean they can be folded on, now?

And Gerbil Jedidiah, when will you get a GTX 480 to start folding on?

And any others who have a 400-series GPU, are you willing to give it a test? It's something you can do one time, and help out our team, Team 2630, while possibly helping medical science if you continue (and even doing it once, if it succeeds, will help).

[Gerbil Jedidiah]

http://techreport.com/discussions.x/18962

According to the above link, nVidia 400-series GPU's now have CUDA 3.1-enabled beta drivers. Does this mean they can be folded on, now?

And Gerbil Jedidiah, when will you get a GTX 480 to start folding on?

And any others who have a 400-series GPU, are you willing to give it a test? It's something you can do one time, and help out our team, Team 2630, while possibly helping medical science if you continue (and even doing it once, if it succeeds, will help).

Long story, but I returned a GTX480 unopened. I now have a Sapphire Vapor Cooled 5870 that runs fast and cool. I don't fold on it though. Right now I just have two GT240s and my i7 920 folding. They have the best points per watts.

[Ragnar Dan]

I don't see much talk on foldingforum.org about the new GPU's, which leads me to think they may not be ready for folding (or at least the folding client isn't ready for that version of CUDA). I think Folding requires the older CUDA version (2.1 or even 2.0, I forget which), so it may remain a waste for anything other than games and for helping with heating or for A/C testing, depending on the time of year and hemisphere.

[Ragnar Dan]

Well, it appears there is a way to use the GTX 400 series, and I was wrong about CUDA versions in use, too, according to this FAQ for a beta of a new core running on the supposedly production GPU3 client: http://folding.stanford.edu/English/FAQ-NVIDIA-GPU3. Apparently I'm running CUDA 2.3, FWIW.

[Ragnar Dan]

Since my Linux SMP client has been acting poorly for the last couple of WU's, running at less than half its normal speed even though it still gets most of the processor, I've decided to stop running it and see if I can get the Windows SMP client to do better. Now that I look at that stuff, though, it looks like they're still using the same silly "you must choose either DEINO or MPICH, and we're not going to give you any indication what the differences are other than one has 64 bit OS support." Great.

[just brew it!]

Since my Linux SMP client has been acting poorly for the last couple of WU's, running at less than half its normal speed even though it still gets most of the processor,

Are you getting A3 WUs, or the A1/A2 WUs?

[Ragnar Dan]

A3's. It was normally claiming from ~260 when I was doing something in Windows that used up processor, to ~320 PPD in A3 WU's on Fahmon, which works out not so great, but roughly equivalent to other older AMD-based machines I run. But it just decided to give me lousy production even though the vmware process would get over 90% of the CPU. I restarted the guest OS, restarted the host OS, and it still ran the same way, ~160 PPD and less. So I'm giving up for the time being, and using the opportunity to test Windows SMP and maybe later a newer Ubuntu than the one I had on there.

[Flying Fox]

A3's. It was normally claiming from ~260 when I was doing something in Windows that used up processor, to ~320 PPD in A3 WU's on Fahmon, which works out not so great, but roughly equivalent to other older AMD-based machines I run. But it just decided to give me lousy production even though the vmware process would get over 90% of the CPU. I restarted the guest OS, restarted the host OS, and it still ran the same way, ~160 PPD and less. So I'm giving up for the time being, and using the opportunity to test Windows SMP and maybe later a newer Ubuntu than the one I had on there.

Could be just the last few WUs that you are encountering. All my Linux VM SMP clients have been doing pretty ok.

Yes, WinSMP still gives you that business so IMO it is not ready for primetime yet. Once they remove the A1 and A2 cores, they can kill this multi-process nonsense and hopefully WinSMP will be much improved to a point we can recommend it.

[just brew it!]

I've seen a slight drop-off in production over the past week or so as well. Probably a couple of wonky WU types in the current mix.

[Evaders99]

There is a beta release for GPU3 client, which should support the Nvidia 400 cards.
http://foldingforum.org/viewtopic.php?f=24&t=14671

[Ragnar Dan]

Could be just the last few WUs that you are encountering. All my Linux VM SMP clients have been doing pretty ok.

Yes, WinSMP still gives you that business so IMO it is not ready for primetime yet. Once they remove the A1 and A2 cores, they can kill this multi-process nonsense and hopefully WinSMP will be much improved to a point we can recommend it.

All my other Linux clients are fine too, and there's nothing odd looking about these WU's I've gotten, so I can't see why they'd perform so poorly. But as of now, my Windows SMP client using MPICH isn't doing all that well either. I hoped it wouldn't need a command window by now, but ignoring that, it has varied from 20.5 minutes in one frame to 44m 50s in another, though I think I began checking email during that last one. Still, they're all taking over 37 minutes except that one frame, and that means it may not complete before the WU expires. And that makes me think there may be something goofy going on. 3 WU's in a row are unlikely to be especially long to complete. And BTW, I tried to move the slow WU from Linux to Windows, but the queue.dat didn't work or something and it gave me an error and promptly deleted the WU and downloaded the current one, so meh.

And Process Explorer provides no evidence of anything stealing cycles that I can find so far anyway. I'm stumped.

[Ragnar Dan]

I figured out one problem my Windows SMP client had. I forgot to add "-forceasm" to the command line for the client, which meant that whenever I killed it, for whatever reason it would restart by saying, "Working with standard loops on this execution." That thing should be in CAPITALIZED BOLD TEXT so it's easier to see.

I never noticed it until looking for something else.

Edit: I see that it doesn't seem to make that big of a difference, oddly enough. I'm going to miss the deadline for the WU, but only by a couple of hours. It makes it absolutely worthless for me to keep running them. If I could get A2 WU's, I'd probably do that instead. My last try for this machine will probably be a virtual Ubuntu 10.04 install, and if that doesn't work then I'll give up and probably replace the system with a C2D in a while, which means I'll have more useless DDR and a mobo of little value, and have to buy some DDR2 which is expensive (or hope my replacement Crucial which I haven't sent for yet works for once). Or I could try to sell the C2D and see if I can get something cheap in the DDR3 world.

[Flying Fox]

I figured out one problem my Windows SMP client had. I forgot to add "-forceasm" to the command line for the client, which meant that whenever I killed it, for whatever reason it would restart by saying, "Working with standard loops on this execution." That thing should be in CAPITALIZED BOLD TEXT so it's easier to see.

I never noticed it until looking for something else.

Edit: I see that it doesn't seem to make that big of a difference, oddly enough. I'm going to miss the deadline for the WU, but only by a couple of hours. It makes it absolutely worthless for me to keep running them. If I could get A2 WU's, I'd probably do that instead. My last try for this machine will probably be a virtual Ubuntu 10.04 install, and if that doesn't work then I'll give up and probably replace the system with a C2D in a while, which means I'll have more useless DDR and a mobo of little value, and have to buy some DDR2 which is expensive (or hope my replacement Crucial which I haven't sent for yet works for once). Or I could try to sell the C2D and see if I can get something cheap in the DDR3 world.

Are you getting A1's or A3's now?

[Ragnar Dan]

All A3's.

I finally did the only test left, which was copying the necessary files to a flash drive to continue the WU it was on, and see how it compared under a 6.29 updated notfred version booted from a CD. It took about 30 minutes per frame after doing a partially completed frame and then doing an entire new frame. Which is odd, because calculating the time per frame from the partial made it something around 23 mins and some seconds per frame, but whatever the case, moving back to a virtual 10.04 Ubuntu LTS install through VMware 1.0.9 it has been taking nearly an hour per frame, which is far too long. Especially since I gave VMware Normal priority.

I'm considering trying it under Windows 7 Pro and seeing how it does there, but I'm dubious. I'd update the VMware version, but last time I tried a newer version it was a mess, used more CPU and harder to get keyboard input to go to the proper OS.

[Ragnar Dan]

I decided to drop the idea of the Windows 7 test. I think my problem is a possible combination of things, and I lack the drive to verify my suspicion for the time being.

The Windows SMP client is just slower, and for whatever reason 2 cores isn't working well with both SMP and the GPU clients in XP Pro. I've never tried the Windows client before on the machine, so I have nothing to compare to, and of course the A3 WU's change things enough that it's probably not a useful comparison anyway.

I think the other problem is that I shouldn't have tried using Ubuntu 10.04 LTS. It behaves as somewhat of a pig from what I can see. Mine can't even get keyboard input right. Typing as I normally do, multiple letters will show up. Trying "ls", for example, will often give me "lllllllss" or something. That kind of crap is more trouble than it's worth. I may try running 9.04 instead if I become ambitious and get a couple of dozen other useful things done first.

[Flying Fox]

I decided to drop the idea of the Windows 7 test. I think my problem is a possible combination of things, and I lack the drive to verify my suspicion for the time being.

The Windows SMP client is just slower, and for whatever reason 2 cores isn't working well with both SMP and the GPU clients in XP Pro. I've never tried the Windows client before on the machine, so I have nothing to compare to, and of course the A3 WU's change things enough that it's probably not a useful comparison anyway.

I think the other problem is that I shouldn't have tried using Ubuntu 10.04 LTS. It behaves as somewhat of a pig from what I can see. Mine can't even get keyboard input right. Typing as I normally do, multiple letters will show up. Trying "ls", for example, will often give me "lllllllss" or something. That kind of crap is more trouble than it's worth. I may try running 9.04 instead if I become ambitious and get a couple of dozen other useful things done first.

Interesting, I think I have the "llllsss" problem when I was using 9.10 with VMware Server version 1. Have you tried a new VMware Server version, and/or install the Dual Core Optimizer if you are using AMD CPUs (it felt like a timing issue to me at the time)?

[Ragnar Dan]

I'm still on version 1.0.9, which is the last one which worked well for me. I also have version 1.0.10-203137, whatever that is, and I can't remember if it is the one which gave me trouble or not any more. I first downloaded it in early November 2009, and then again in late December, so maybe they have a newer version since then. I tend to stick with something if it works, and lately I'm not having much luck that way.

Edit: I found I have a newer version, 2.0.2, but that awaits frustration before use.

As for a dual core optimizer, if you mean for Windows, I never needed it before, though I have it. The one for XP was originally released in 2006, and I'm fairly confident all MS OSes have already accounted for any such necessary changes.

[Flying Fox]

I would suggest VMware Player at this point then, unless you need remote management capabilities. You can assign more than 2 CPUs with v3.

[Ragnar Dan]

Hm. I tried the AMD dual core optimizer I had sitting among my software updates ~5.5 hours ago, since I had it after all, and though I thought I never used it, it's been there since 2006 so I could have forgotten about it by now. And from the results that does seem plausible. I'm well past the deadline for the WU, but it's been doing ~23 minute frames and I can even give it BelowNormal priority now, and it continues at the same general rate depending on foreground activity. Which is to say, OOPS!

Thanks for reminding me about that thing.

If I did use it before, though, it must have been after I started doing the VMware->Linux SMP thing, because I specifically recall wondering what people were seeing that made them use it when I first learned of it and deemed it unnecessary in late 2005 when I built that machine. I think it was Winter (Jan./Feb?) 2007 when I began the SMP folding bit after Hotdog posted that thread about "Hotdog's experiences" in this forum.

I still can't see why the newer Linux version lets me move the pointer outside of the VM window without having to send the escape characters (CTRL-ALT) to do it like my older Ubuntu install did, even with the current version of VMware. But it's not that big of a deal.

[Flying Fox]

Hm. I tried the AMD dual core optimizer I had sitting among my software updates ~5.5 hours ago, since I had it after all, and though I thought I never used it, it's been there since 2006 so I could have forgotten about it by now. And from the results that does seem plausible. I'm well past the deadline for the WU, but it's been doing ~23 minute frames and I can even give it BelowNormal priority now, and it continues at the same general rate depending on foreground activity. Which is to say, OOPS!

Thanks for reminding me about that thing.

If I did use it before, though, it must have been after I started doing the VMware->Linux SMP thing, because I specifically recall wondering what people were seeing that made them use it when I first learned of it and deemed it unnecessary in late 2005 when I built that machine. I think it was Winter (Jan./Feb?) 2007 when I began the SMP folding bit after Hotdog posted that thread about "Hotdog's experiences" in this forum.

When I made the jump to an X2, not installing the dual core optimizer gave me negative pings and all sorts of ****. VMware in particular is pretty bad too because the timings are all wacked.

BTW, where is Hotdog? We missed his points.

I still can't see why the newer Linux version lets me move the pointer outside of the VM window without having to send the escape characters (CTRL-ALT) to do it like my older Ubuntu install did, even with the current version of VMware. But it's not that big of a deal.

It has been there since 9.10 I think. I am always under the impression that they put something in the kernel/drivers to support VMware out of the box or something. Save us the trouble of installing VMware tools. You really should use the newer Server v2 or Player v3. Server v2 is pretty retarded usage wise, but it is still pretty solid piece of software once you look past the usability problems. Player v3 actually kind of looks like VirtualBox (or VB copied VMware Player?), and it supports more than 2 cores.

[Ragnar Dan]

When I made the jump to an X2, not installing the dual core optimizer gave me negative pings and all sorts of ****. VMware in particular is pretty bad too because the timings are all wacked.

That's pretty odd. I suppose the timing will do that, but negative? Sheesh. ICMP says, "I knew you were going to ping that site, so I started before you asked."

BTW, where is Hotdog? We missed his points.

I dunno. I think when he quit there was talk about it, but I can't remember the result.

I still can't see why the newer Linux version lets me move the pointer outside of the VM window without having to send the escape characters (CTRL-ALT) to do it like my older Ubuntu install did, even with the current version of VMware. But it's not that big of a deal.

It has been there since 9.10 I think. I am always under the impression that they put something in the kernel/drivers to support VMware out of the box or something. Save us the trouble of installing VMware tools. You really should use the newer Server v2 or Player v3. Server v2 is pretty retarded usage wise, but it is still pretty solid piece of software once you look past the usability problems. Player v3 actually kind of looks like VirtualBox (or VB copied VMware Player?), and it supports more than 2 cores.

Yeah, I thought it was 9.04, but maybe I had updated by then. My 10.04 doesn't seem to be able to detect my sound hardware for some reason, unlike all earlier versions.

As for your sales technique, it needs improvement. The VM's usage/user interface is the main thing that distinguishes it for me since they started allowing more than 1 core. I have only 1 quad machine, and until I upgrade it's all I'll have, and is a dedicated notfred box. If I could get ssl on notfred's stuff, I'd be happy, though.

[just brew it!]

That's pretty odd. I suppose the timing will do that, but negative? Sheesh. ICMP says, "I knew you were going to ping that site, so I started before you asked."

The underlying issue is that there's a hardware timer integrated into the CPU core. Some software (including ping, apparently) uses that timer to measure short intervals of elapsed time. Unfortunately, some of AMD's multi-core CPUs did not guarantee that the hardware timer was synchronized between multiple cores in the same CPU. So if a process got rescheduled from one core to another, time could appear to run backwards.

The "dual core optimizer" isn't so much an optimization, as a hack to make sure that time never runs backwards when processes are shuffled between different cores.

IIRC newer AMD CPUs (Phenom?) fix this issue in hardware.

[Ragnar Dan]

Yeah yeah, stop trying to kill my small fun-making.

But earlier AMD processors must have had a rather huge problem with processes being bounced around to make that big of a difference in production on my VMwared client. Which is what I'm getting out of it. The main thing I've done pings on since moving to "cable net" is testing connectivity, not so much speeds, so I can't be sure I never would have had negative results, but I never saw one anyway.

[farmpuma]

I've been trying for over an hour and the client tells me no server can be found!

The 171.64.65.54 server listings are missing from the Project Summary.

And the listing in the Server Status is all zeroed out!

[Ragnar Dan]

Yep, and I've got 2 uploaded A3 WU's which haven't been credited yet, either.

Edit: Here's the thread with the error reported by Kasson: servers 171.64.65.53 and 171.64.65.54 down.

[farmpuma]

Yesterday I eventually received a 6023 from 171.64.65.54 which I know from previous experience violates my configuration settings. It processed for a minute or two and then hard locked the system. Deleted defective WU.

Later I received a 6025 from 171.64.65.54 which I know from previous experience violates my configuration settings. It is processing without other issues, so far.

BTW

My experience seems to indicate that the A3's should give a marginal improvement to the faster A1/A2 WUs

My experience based on actual bonus points awarded shows the online calculator linked by Star_Platinum to be quite accurate. It also shows the penalty of the 'bonus system' to be about ten to fifteen percent compared to the good A1 work units. But considering how badly some of the A1 WUs stank, it's probably an overall gain in average PPD. Provided one doesn't receive any of the other penalties involved with the 'bonus system'.

[Flying Fox]

My experience seems to indicate that the A3's should give a marginal improvement to the faster A1/A2 WUs

My experience based on actual bonus points awarded shows the online calculator linked by Star_Platinum to be quite accurate. It also shows the penalty of the 'bonus system' to be about ten to fifteen percent compared to the good A1 work units. But considering how badly some of the A1 WUs stank, it's probably an overall gain in average PPD. Provided one doesn't receive any of the other penalties involved with the 'bonus system'.

I would tend to think HFM.net's numbers would be more accurate since it stores previous results for better benchmarking. But in the end you know there will be discrepancies as who knows what Stanford is doing.

[farmpuma]

I setup my passkey as farmpuma@a3. Because @ is an invalid character Stanford doesn't display it or anything after it, but still lists it separately from all the other farmpumas arranged by the characters after the @. Since the difference is invisible to EOC I know exactly what Stanford is doing and still maintain the correct farmpuma total with EOC stats.

[Ragnar Dan]

All my machines have been getting those 921 point P6701 WU's lately, except my Opty 165 which got an A1 core P5102 (3340 points).

The problem: my old X2-3800 running at stock, now, won't make the deadline. If I'd ever had success transferring SMP WU's from one machine to another and back again, I might consider moving it to another machine temporarily and then moving it back for the final 1% and upload, but I've usually gotten a checkpoint error and a restart from 0% after 2 transfers. Bah. I could quickly buy a mobo for my old C2D 6400 and move it to that thing, and it might make it. But I'd have to get the thing within the next 2 days, and even then it might not make it (as of this typing under 55⅔ hours until bonus runs out and the WU is re-issued by Stanford, and after less than 21.25 hours it's already ~3⅓ hours behind).