What’s the optimal Folding client combo?

Team TR’s Folding production has been moving upward since our recent call to arms, I’m happy to report.

But, I must sheepishly admit that I haven’t been Folding on my main PC since building myself a new system. I’d like to install the right clients on my PC and schedule them to run when I’m not around, but I have an honest question that some other folks seem to be asking, too. Here it is:

My PC has a multi-core (quad Core 2, in my case) processor and a graphics card (Radeon HD 3850 512MB) capable of running the Folding@Home GPU client. What mix of clients should I be installing to get the most points out of this box? One instance of the SMP client and one of the GPU client? Two SMP and one GPU? Or is that too much?

How would things change if I had a dual-core CPU?

Also, if I want to schedule the clients to start and stop a certain times of the day, does that complicate matters?

We need to have a clear sense of what folks should be running so we can encourage best practices and max points. Team leaders and other vets, what do you recommend? Any installation tips to make the best combos work together smoothly? Thanks.

Comments closed
    • BooTs
    • 11 years ago

    I think it’s important to say that the little bits help!

    My home PC has been crazy unstable for a long time and I just found out that it was bad NIC drivers all along. I’ll have the F@H client back on there after this weekend, and if I get around to building a newer PC that will have multiple instances of F@H running as well.

    • tfp
    • 11 years ago

    If the clients are buggy and crashing, is the science they are producing any good? Unstable things can produce funky results.

      • Convert
      • 11 years ago

      The GUI and console versions for Windows work just fine.

      Furthermore, our qualms are that the client stops running. The client does checks on its own and while there is always a possibility for error it wouldn’t be much different from their current clients all things considered.

      Not running =! corrupt results, or however the code monkeys say it.

        • tfp
        • 11 years ago

        The reason I ask is because in general crash can == corrupted data. I would think they would have tested enough to make sure that doesn’t happen, however I would have thought you test enough to release a usable client as well.

          • Convert
          • 11 years ago

          I know what you meant; it’s just that it really doesn’t apply here, thankfully.

          With folding it will revert back to the last checkpoint after a crash, or if it’s really bad it will just abort and send the current results back to Stanford to salvage if at all possible.

          They appear to have focused on getting reliable results as opposed to getting the results reliably, if that makes any sense.

            • farmpuma
            • 11 years ago

            The SMP client deletes the WU when an error occurs and sends back nothing. Luckily I’ve only had it happen a couple of times in over 300 WUs.

            • BooTs
            • 11 years ago

            I would assume they don’t send each workload to only 1 client. They probably send it to many clients, and if any were to differ from the majority, they would be tossed out.

          • just brew it!
          • 11 years ago

          When the client crashes and is (manually) restarted, it reverts to the most recent checkpoint. My understanding is that they also run each work unit more than once (on different systems), to cross-check the results.

          They’re generally pretty good about labeling the unstable clients as beta releases, and/or restricting the WUs that may cause hangs/crashes to people who have specified the -advmethods option. They need to test these new clients somehow; the best way to test them on the widest possible variety of hardware is to encourage people to “opt in”, by offering a point bonus for being a guinea pig.

          I really don’t see why people make such a big deal about it. If you want stability, run the “tried and true” single-core clients. If you’re willing to babysit your F@h system(s) for some additional points, run the beta and/or -advmethods stuff.

            • Flying Fox
            • 11 years ago

            I agree, but the situation is indeed a bit different. I clearly remember -advmethods and the QMD’s giving a sizeable advantage to “regular” WUs and people were complaining too. Let’s face it, the majority of people get in on Folding for the points, that’s why the system is in place to attract more people. People who do this for the science /[

    • Jon
    • 11 years ago

    Does anyone know if the GPU2 client supports nvidia SLI? I acquired another 8800GTX today and I’d like to test it out in my main rig.

      • Flying Fox
      • 11 years ago

      Last I checked I think you need to turn off SLI and then you can use 2xGPU2 clients.

    • Gerbil Jedidiah
    • 11 years ago

    I’m willing to bet money I have the most highly efficient PPD setup on my computer. I am running Vista64 on a Q9450 system with an 8800GTS G92 videocard. CPU is running at 3.5GHZ. Here’s the setup:

    SMP Client on linux in VM with affinity set to cores 0 and 1

    SMP Client on linux in VM with affinity set to cores 2 and 3

    GPU2 client running on 8800GTS with no affinity set.

    The GPU2 client doesn’t use that much CPU power in Vista 64, so it’s no problem to have each SMP client set for two cores, and then have the GPU2 client just grab a few cycles from any of the 4 cores. I’m getting around 5,000 ppd on my videocard and 2,400 ppd from each SMP client. Here’s a link:

    §[<http://img401.imageshack.us/img401/8215/freshyf5.png<]§

      • Nitrodist
      • 11 years ago

      I have my setup as follows.

      Q9300 (2.5ghz)
      XP Pro 64-bit
      8800GT 512

      Set 2 VM Ubuntu installations up and put the SMP clients on those. Then I set the affinity of the first three cores (Core 0, 1, 2) to those 2 clients (3 cores over 2 clients). The last core (core 3) has its affinity set by default for the GPU client.

      Works great so far. Banging out (sub) 15 minute frames for the SMP clients and getting ~3500PPD from the GPU client. Less PPD for the GPU because it’s a 64bit OS… but I don’t feel like using Vista yet.

    • Tarx
    • 11 years ago

    For a new folder with XP or Vista with an ATI HD2000 series card or higher (and Nvidia 8000 series card and higher), the GPU2 client at §[<http://www.stanford.edu/group/pandegroup/folding/release/Folding@home-Win32-GPU-systray-620.msi<]§ is simple to install and works great. I've done a simple install guide at §[<http://forums.ncix.com/forums/topic.php?id=1763746<]§ (note: the team number for Tech Report is 2630). Although the ATI client is a bit more of a pig in CPU use (idle use) plus gives quite a bit less points for a similar class Nvidia card, it still does relatively well plus it is a bit more polished (as it was the original GPU2 client). For anyone with Linux, the SMP client has worked great for a long time. That's part of the reason why on XP/Vista running VMware with Linux client remains so popular (see §[<http://www.techreport.com/forums/viewtopic.php?f=9&t=47827<]§ for details on setting this up). Don't need to know anything about Linux or VMware to do it (and don't need to learn anything either), plus it's a nice way to be introduced to VMware and Linux (if interested). BTW a GPU2 client on Linux is in the works plus there has been progress getting it to run in Wine. §[<http://foldingforum.org/viewtopic.php?f=52&t=3744<]§ I've given up on the latest Windows SMP client for now (I was suppose to write a FAQ on how to install it, but couldn't get it working - argh!) (BTW the older version did work fine for me) but the Folding@home project leader says that getting this client (and a couple of other PITA clients) to be user friendly is their main priority for improving the clients. I'm waiting until that happens. §[<http://folding.typepad.com/news/2008/08/update-on-winsmp-client-development.html<]§ So my suggestion for new folders on Windows XP/Vista (and those that don't want to setup VMware and not willing to attempt to install the Windows SMP client) is to first install the GPU2 client (if they have a capable graphics card) and then (if up for a bit more fiddly work) install one regular 6.20 console clients for each of the other cores - it pretty similar to what is mentioned in the instructions (instructions do need to be refreshed a bit - farmpuma?) §[<https://techreport.com/forums/viewtopic.php?f=9&t=52890<]§ but if there are any questions, just ask in the forum §[<https://techreport.com/forums/viewforum.php?f=9<]§

      • farmpuma
      • 11 years ago

      Trax wrote, “(instructions do need to be refreshed a bit – farmpuma?)”

      At the moment I wouldn’t be much help as I still run only the 5.04 and 4.0 single core clients. And I already had to much on my plate before the massive thorn tree tried to eat my 40×40 foot grain storage shed. Perhaps I’ll get bored this winter and see if the version 6 widgets can track my UGN donations.

    • just brew it!
    • 11 years ago

    Yes, unfortunately there is a tradeoff between performance and ease of installation/use.

    For simplicity and trouble-free “fire ‘n’ forget” operation, the standard non-SMP console client is the simplest. Just install it once for each core; put each instance in a separate folder, and make sure you give each one its own machine ID when you configure it. This client can operate as a system service, so you can have it fire up automatically when the system boots.

    GPU client is a good option if you have a GPU that is capable of running it. If you’re also running separate non-SMP console clients for each core, decrease the number of console clients by 1 so that one core is available to manage the GPU client.

    I’ve apparently had less trouble with the Windows SMP client than some other people here, but yeah it still has some warts. The recommendation to run the Linux version in a VM is largely a response to the historical instability of the Windows SMP client (later versions are better, but it is still considered to be in beta).

    The Linux SMP client is stable if you don’t use the -advmethods option, but requires that you run a 64-bit distro. Given that all modern x86 CPUs are 64-bit capable, there’s really no reason not to run a 64-bit distro anyway. (And yes, believe it or not you *[

      • notfred
      • 11 years ago

      I have to agree with JBI here. If you want simple install and forget, single core client is it. If you want to get massive points, Linux SMP in a VM plus GPU client is it with associated complexity of install and need for monitoring / regular fixing.

      It’s no more different than computers: you can put together a system at stock speeds, turn it on and know it will just work; or you can spend a bunch of time modding and tuning and squeezing the last ounce of overclocked performance out of it.

      Of course if you combine the above and overclock an SMP/GPU setup, it produces even more points.

      • Tarx
      • 11 years ago

      64bit VM on a 32bit OS needs VT for Intel CPUs. That means the E1000, E2000, E4000, E5000, E7000 and Q8000 series will not work!
      BTW any X2, X3, X4 will work, even if they don’t have VT (only the older X2s don’t have VT) as it has the additional necessary instructions.

        • just brew it!
        • 11 years ago

        Thanks for the correction!

        Turns out I’ve only tried the 64-on-32 trick with AMD X2s, so I never hit the limitation on doing this with older Intel 64-bit CPUs.

    • nerdrage
    • 11 years ago

    This might be strikingly obvious, but casual folders want something simple — install it once, set it, and forget it. There should be one client. It should look at your hardware when it initializes and run in the most optimal way possible by default, automatically. It should allow configuration for those who want to. All configuration changes should be done through a GUI and written to a plain text configuration file, not command-line parameters or the registry. The plain text file should work correctly even if the CR/LF format isn’t exactly right. The configuration options should be uniform across operating systems, and should allow me to limit the resources the client is using, and easily start and stop the client without messing around with services or the command line. It should automatically update itself if I set it to. I shouldn’t have to sift through forum posts to figure out the best way to run it. I shouldn’t have to spend any time “maintaining” it. I shouldn’t have to worry about getting work done on time if I don’t fold 24/7 at 100% CPU utilization with the latest quad-core CPU. It should just sit in the background, do its thing, and not interfere with the real work being done on the PC.

    I’m not asking for anything revolutionary. I don’t think it’s particularly difficult to do any of this. Why is it so hard for Pande group to get it?

      • Flying Fox
      • 11 years ago

      Believe me everyone wants all that too, and I’m not defending the Pande Group here. I think they are responsible as well, but with their small team I would rather them focusing on the science and publishing more results first if it means software improvements come more slowly. Of course if they have more resources they can get some real software developers (heck Stanford is supposed to have tons of good programmers in their CS/EE program right?) to really fix these things up. So that’s a little puzzling that nobody is attracted to such a high profile project.

        • nerdrage
        • 11 years ago

        But it’s been YEARS like this, with hardly any improvement. It has the feel of a high-school CS project. I’m only a junior software developer — but most of the things I mentioned shouldn’t take much time at all to implement. Ideally, it could be split into a front-end/back-end model. Pande could do the back-end scientific stuff, and let someone else (with software skills from the current millennium) write the front-end.

          • Flying Fox
          • 11 years ago

          Somehow they seem unwilling to do that. If they had opened up the interface to the backend I’m sure we will have hundreds if not thousands of frontends already.

          The single core client took years to develop. We only just had the SMP for a year or so, GPU2 is only a few months. Need to give them some time.

      • eitje
      • 11 years ago

      Rolling your own “FAH Detector” app wouldn’t be too difficult, right?

      Imagine a single window app. when it’s launched, it gets your hardware configuration and checks for various installed hardware components. At the end, it either automatically redirects to the “best” installer for your system, or just lists the installers that are “compatible” with your system.

      I’m a .NET guy, so I’d probably do it with WMI:
      §[<http://msdn.microsoft.com/en-us/library/system.management.aspx<]§ §[<http://msdn.microsoft.com/en-us/library/aa394554(VS.85).aspx<]§ Here are some examples: §[<http://msdn.microsoft.com/en-us/magazine/cc302051.aspx<]§ §[<http://www.codeproject.com/KB/vbscript/VvvHardwareInfo.aspx<]§ §[<http://msdn.microsoft.com/en-us/library/aa394587(VS.85).aspx<]§ That actually sounds like a fun project. 🙂 And, although it doesn't solve everything you're looking to do, it definitely takes care of the first part of your post. And Damage's question, actually.

        • nerdrage
        • 11 years ago

        I’d take it one step further though — don’t detect the hardware during installation — do it each time the client starts up. That way it will automatically take advantage of new hardware when you upgrade your video card or CPU. And the rules about what’s “optimal” can be part of the automatic client updates.

        Maybe you can volunteer with Pande and get them in shape!

        I know, I know… wishful thinking.

          • eitje
          • 11 years ago

          You can always identify the dreamers in a conversation. 🙂

          I’m talking about something that you could do now & immediately to contribute to the community as a whole, even as a junior software developer.

          You’re talking about something which I would need to engage the Pande team to accomplish, and it’s been proven time and again that the Pande team has no interest in working with others when it comes to their software.

          One of these things is possible and pragmatic, the other is not. 😉

    • Lucky Jack Aubrey
    • 11 years ago

    I have had very good luck running the GPU client plus two standard CPU clients on a dual-core machine. I was cranking out around 5K points a day on just one machine until I ran into a hardware problem unrelated to folding (I will have it running at full tilt again soon).

    • eitje
    • 11 years ago

    After reading the back & forth discussions here in this thread (the same ones that I remember from the past few years, in fact), let me turn your question back around on you, Damage:

    Where are the benchmarks that would really answer this question? Perhaps one of the new TechReport Minions could look at writing an article, with benchmarks, addressing the different combinations and the resultant PPD that each would theoretically offer. All said Minion would need is a list of combinations to try, and some time.

      • Flying Fox
      • 11 years ago

      Issues with that:
      – the client is not super stable yet, mpiexec still sucks on Windows
      – the WU values are not fixed yet due to the beta nature of the client
      – the ppd values for different WU types vary quite a bit, so if during your test run you get a 2605 vs a 2665 you get very different WU values, I’m seeing pretty sizable variants for GPU2 WUs as well

        • eitje
        • 11 years ago

        so, you’re saying, we can’t run any tests because the projects work units & values are being mismanaged?

        I think I would rely upon averages, and several days of running in each client configuration. If the SMP client requires babysitting, then that should be brought up in the article, and reflected in the numbers it generates.

          • Flying Fox
          • 11 years ago

          They are not mismanaged, it is just the betaness that makes everything subject to change rather quickly.

          It’s the unreliability and inconsistencies that are hampering testing, not to mention we need notfred’s time to do the benchmarking tool, and he’s busy with just keeping up with the new betas and the other features of the diskless stuff.

          For now, we can rely on field reports provided the person(s) does the measurement semi-properly.

            • eitje
            • 11 years ago

            you don’t need notfred’s time.

            the tester takes his system, images it, and installs configuration X. He then lets it run for N work units (let’s say N=6). At the end of that time, he checks whatever stats he’s benchmarking: PPD, power used, whatever.

            Then, he reimages the system and installs configuration X+1. At the end of 3 weeks, he’s tested some number of configurations, and can provide empirical data for each one.

            • Flying Fox
            • 11 years ago

            The problem is the set of N work units may be different due to WU assignments. Let’s say N=4. You may get a couple of 2605’s and a couple of 2665’s, or all 2605’s or all 2665’s. The ppd number of those sets are vastly different. 🙁

            The single core WUs are pretty consistent ppd-wise with different categories between Ambers (the old ones), Gromacs, etc. However, all current SMP WUs are GROMACS-SMP and they vary so much in ppd. So you can keep fishing 2605 or 2665, but I think it is too much work.

            • eitje
            • 11 years ago

            I think N would need to be a sufficiently large number, such that statistics could come into play. We could figure out what that number is, if we knew the count of unique WUs available for any given client.

            Also, I believe we’ve hit on the fundamental difference between you and me – I don’t think it’s too much work, since it’s just setting up the client and letting it run until a certain number of WUs have completed.

            Frankly, if I had the money to buy all of the necessary hardware, I’d just start running the tests myself. But that looks like it could cost about $600-800, to get the video cards, quad core, motherboard + RAM that would be necessary.

            That’s not an issue for someone that already has the parts, though, and they’d be contributing to TR while they benched. 😀

            • Flying Fox
            • 11 years ago

            You basically just hit the jist of the problem. I would rather Damage and the TR staff go to IDF, Nvision08, benchmark other cool hardware, deal with the families, and other things before committing weeks if not months of their time doing this. Which means someone on our team need to start reporting on the data and someone else needs to help draw some conclusions. Damage Labs may contain a lot of hardware, but I am not sure if they can afford to run a farm with cutting edge hardware all clogged up in running N work units when those hardware can be used for other testing.

            • eitje
            • 11 years ago

            then go organize your team and get cracking! 🙂

            • Flying Fox
            • 11 years ago

            I thought you are part of team 2630 too, no?

            • farmpuma
            • 11 years ago

            Ok ok, I’ve been meaning to address this issue for a while. There is a simple solution that doesn’t require an infinite number of WUs or an infinite amount of lab time for the SMP clients.

            As notfred did with his benchmark cd all one really needs is a random snapshot of a random SMP WU. Just copy the WU work folder, queue.dat file, and unitinfo.txt file (for easy identification) into a temp (SMP_test_WU?) folder. Download an SMP client and tell it to ask before connecting during the initial launch configuration. When it asks to connect close the client (control-c). Delete the newly created work folder and the queue.dat file. Copy your test WU folder contents into the client folder and restart. Allow the client to download the appropriate core and fold for three or four percent before closing the client. With FahMon configured to calculate PPD for the “Last 3 frames” you don’t even need to crunch the numbers in the logfile. Rinse and repeat for different hardware or SMP clients or install variations.

            edit2: Note! The system clock of the hardware under test must be set to a time and date within the deadline range of the “test WU”. This range is listed in the unitinfo.txt file.

            It might be a good idea to use a 26xx, a 30xx, and a core a2 “test WU” to give a more complete overview, although one would expect them to scale in a similar fashion.

            edit: To the original topic, I would recommend one SMP client in a VM as it defaults to two cores (or set core affinity to only two cores) and a GPU client. This should give you a free core and change (depending on your OS) to keep us informed and on track.

            • eitje
            • 11 years ago

            that approach could probably even be scripted, such that user input wasn’t required.

            If I wasn’t working on a DBP project outside of work right now, I’d probably get on that. I’ll have to make a note to myself, somewhere.

    • JPinTO
    • 11 years ago

    Assuming you can overclock the quad to 3.0Ghz, you should be able to get 4000-4500PPD with a pair of Linux SMP clients… 1 client per 2 cores.

    You can add a GPU client , but you will give up 25% production from the SMP guests so the 3850 better be doing more than 1000PPD. If it doesn’t then buy a 8800GT for $100 which is good for 5000PPD on it’s own + the ~3000 from the 2 SMP clients.

    Fold on!

    • SmokinJoe-Salem
    • 11 years ago

    Everything changes so fast this is a tough question. After 8+ months folding with the SMP windows client (Q6600, Vista-32), the VMware, Ubuntu, Linux-64bit is simpy a better way to fold. Trax’s install guide is excellent, and you don’t need to know much about WMware, Ubuntu, or Linux to get it up and running. The Windows SMP client and the GPU client don’t work well together.

    • SuperSpy
    • 11 years ago

    I just installed the GPU client on my HP dv9500t (dual core C2D 2.0GHz, GF8600M GS) and so far I have been pretty impressed. CPU usage is pretty much nil, while the GPU appears to be doing quite well for itself. From my basic reading of the tooltip, it appears to be doing 1% (250/25000) every couple minutes or so.

    If I’m feeling adventurous, I might try the SMP client on top of this, but I fear that might be a bit much for a laptop. =)

    • Pegasus
    • 11 years ago

    I used to run the Windows SMP client. But it would often hang and I had to restart the client too often. That was my experience on both an AMD dual and an Intel quad box.

    Right now I run 2 VMware linux installs on my quad. It never hangs. Always is running. It was pretty easy going through the guide in TR’s forums. Even though it was old it was easy figuring out the right download path etc. And I can let my ATI 4850 also run FAH by telling the VMware setups not to use the last core. So I run 2 SMP clients over 3 cores. For each SMP, 1300ppd normal and 1800ppd with -advmethods and a A2 core WU. Q9300 stock.
    Just had to put in the .vmx file for each setup the following.
    processor3.use = “FALSE”

    • onlycodered
    • 11 years ago

    Actually, the ATI GPU client isn’t really worth it at the moment since it’s somewhat limited. I believe Stanford is working on this as quickly as possible, but it may take a while.

    • Convert
    • 11 years ago

    Don’t bother with the SMP client, it’s a complete POS. Feel free to give it a spin if you have plenty of time to completely remove it and reinstall it every couple of weeks.

    I would just run multiple instances of the standard client. No fuss, no worries, it just works.

    The GPU client is also going to give you a few issues but it’s not as bad as the SMP. The PPD on it though really does outweigh the headaches. My entire farm has been down for the last month? because all of my SMP clients stopped working (which will require a full reinstall of the client) and the GPU client also stopped, which will require the same ritualistic rain dance to get it loaded back on and functioning again. I have been so frustrated with the POS programs I am not looking forward to doing it again for the 50th time. Maybe I will get smart here and move the farm back over to non smp/high production clients.

    I don’t want to discourage anyone, please don’t take it this way, just a word of warning and hopefully this will set realistic expectations and no one will get frustrated when everything craps out and decide to quit folding outright. The standard clients are rock solid so those of you hesitant on having to invest much time just run that version and it will be worry free.

      • Damage
      • 11 years ago

      No wonder Team TR has been having trouble. Ugh. The client situation is just grim.

        • derFunkenstein
        • 11 years ago

        Maybe it’s time to find a new DC app. Berkeley’s BOINC has a Folding@Home client (separate from Stanford’s) and it’s much better-supported.

        • Usacomp2k3
        • 11 years ago

        It is quite grim. That’s why I’ve stopped folding myself (I posted a rant somewhere to that effect). I’m happy to support those that can handle the crappy solutions though.

          • Convert
          • 11 years ago

          You had problems with the console version (non SMP)?

        • Flying Fox
        • 11 years ago

        To me this is not so different from when the single client was still in its infant stages. Heard lots of people complaining about it and I stayed with UnitedDevices instead.

        The problem now is that the difference for not using those “high performance” clients is very big now in terms of points, and it shows.

      • gyrfalcon1
      • 11 years ago

      Yea. I’ve tried to get the SMP client running many times on my dual core box. After having it fail, crash, and basically explode multiple times I’ve been forced to just use the standard client which puts out pitiful PPD in comparison 🙁

      • Flying Fox
      • 11 years ago

      I don’t know about you, but I don’t seem to have much problem with both SMP and the GPU2 clients. Most of the time I ran the SMP client in a VM behind a software router. I used notfred’s diskless stuff and haven’t had problems for more than 6 months. The only time it needs maintenance is when I shutdown my host machine and the beta expired. So I just need to regenerate the folding cd iso and restart the whole setup. As for WinSMP, I found that if I have to reboot my router or if I am on a wireless connection and it breaks, the client will hang. If I catch it soon enough, I can just Ctrl-C the thing and it will restart off the last checkpoint. During my trip back in May, I was away for a few weeks and both the VM Linux diskless SMP and WinSMP clients did not die on me.

      After I came back from the trip I installed the GPU2 client on my HD3450. The ppd sucks (still better than the single core client), but it doesn’t seem to have issues for me.

      But I do agree, the new clients are not exactly “set it and forget it” for /[

        • Convert
        • 11 years ago

        I run the WinSMP client. Perhaps there is an easier/more reliable way to go about it but running a VM is certainly not the path I am interested in taking. I have given the notfred diskless stuff some thought but I still need functioning systems.

        I don’t think the high performance clients are set it and forget it for anyone on the planet personally, then again I want to start poking eyes out so maybe I am not giving it much chance.

          • Flying Fox
          • 11 years ago

          You may be running into drfish’s scenario where the network connectivity is not stable enough and you get hangs and stuff. At home my WinSMP ran for 3+ weeks without a hitch, until I decided to power cycle my router. Meanwhile the VM SMP client continues to chug along since I have a separately NAT’ed subnet with a software router (Win2K RRAS), so the VM never loses the IP.

            • Convert
            • 11 years ago

            Why would network connectivity matter? It only uses the network when it sends results?

            The folding machines are plugged into a switch that has had perfect uptime. While my modem and router occasionally need a power cycle they wouldn’t have a direct impact on the network status on any of the folding clients.

            To be clear I run xp as the OS on a dedicated machine with the WinSMP client loaded.

            • Flying Fox
            • 11 years ago

            Yeah it sounds like the problem that I encountered on my WinSMP machine. Whenever you restart the router, Windows will get that “network disconnected” message and for a moment your IP will become 0.0.0.0. The WinSMP client uses the Windows port of mpiexec.exe which uses some sort of local network protocol to communicate with the cores. People have been whining about the poor implementation of mpiexec especially on Windows. You can find discussions in our forum and the official one.

            When the problem hits, I think (can’t remember) the client will continue to work until it hits 100%, then it will just hang there. If you restart the client soon enough (before the deadline?) though, then it will be fine.

            That WinSMP box is now moved to be connected to my LAN wirelessly. So I have switched to single core + GPU2 (on a crappy HD3450) because I know the frequent disconnects will give WinSMP problems to no end. Don’t worry though this is temporary I should be back on WinSMP just in time for our annual Frankie run.

            • Convert
            • 11 years ago

            In my case though they are connected to a switch, the network is not interrupted if the router or modem gets power cycled.

            • Flying Fox
            • 11 years ago

            And I am not disagreeing with you that the SMP client has problems. I was whining about it too. I just don’t think it is a complete POS, yet. notfred has to implement a “hang check” to auto-restart the client for his diskless stuff too.

      • Krogoth
      • 11 years ago

      The latest beta of SMP client is made of fail. It is even more lame that you have to put in either -smp or -smp -denio option for it to work properly.

        • Flying Fox
        • 11 years ago

        The -smp switch is for the future where they merge the SMP and single core clients, which they have already done on the Linux side. The -denio switch is for using new optimizations with MPICH to gain some more efficiencies. This is typical of beta software to try new things.

        Besides, you can add those switches when you are doing -config/-configonly so you don’t have to type them all the time.

    • Geatian
    • 11 years ago

    I agree with Damage; I don’t really see the point in running VMs for a few extra ppd. It’s a lot of hassle for very little payoff. I just run the windows cmd line clients. I have a quad core 2 cpu, and a 4850 in my box, so I run a single SMP client, and the GPU2 client. I wrote a little batch file to start them up:
    q[

    • flip-mode
    • 11 years ago

    Pikaporeon, just an outsider’s perspective….

    If you’re looking to get new people to fold, I’m not so sure the multiple virtual linux machine route is the best way to slowly wade into the folding waters.

    Just from the few posts here, this is what it sounds like the state of things is:

    If Geforce 8, then GPU client
    Else If Geforce 8 and quad, then GPU client and SMP client
    Else SMP client

      • Flying Fox
      • 11 years ago

      That’s exactly right. Only once you get people to be addicted and are hungry for more points, then you introduce them to the “secret to get more points”.

    • Damage
    • 11 years ago

    Let me ask this, then. The whole VM issue seems like a pain to me. Is it really necessary still? I doubt we’ll recruit a lot of folks to go through lots of setup pain, and I don’t have loads of free time myself. What should people who simply wish to spend 15 minutes installing a client or three do?

      • Flying Fox
      • 11 years ago

      Then grab notfred’s VMware appliance and install the free VMware Player?

      A more appropriate answer:

      For the casual folder, you can probably just install one instance (remember to have .NET 2.0 and run that install.bat with a proper user account, this can still be a pain for newbies 🙁 ) of WinSMP and then one GPU2 client.

      If you want to chase after points, then you should look at those crazy VMs and other stuff (including messing with affinity) to /[

      • JPinTO
      • 11 years ago

      Just run the GPU client then… it is stable.

      • Nelliesboo
      • 11 years ago

      After reading post after post about this from I what see is…

      SMP client is just to much work because you have to watch it all time. So what everyone should do is install the old school client as many times as you have a CPU core (making sure each client is set to it’s own cpu core?). Then walk away. As for the GPU client if you want to get a little more fancy Install this if you dont mind checking on it once or twice a week (note you would have to turn off one of the cpu core clients to run the gpu client) Does that sound like right to everyone?

    • leor
    • 11 years ago

    I have an 8 core system, anbd right now I’m on 2SMP and 1 GPU (8800GTX) are you guys suggesting I should have a 3rd SMP running?

      • Flying Fox
      • 11 years ago

      The SMP client is designed to run optimally on 4 cores (you see 4xFahCore_a*.exe running). So 2xSMP+GPU2 should be pretty good. But of course you can test a bit to see if 3xSMP will give you more overall points. Then it becomes similar to the old HT situation, where it becomes more points vs faster results.

      • JPinTO
      • 11 years ago

      For Windows: 4 cores per client is correct, but the PPD is less than stellar.

      For maximum PPD then 4 Linux SMP guests running on vmware is the best way to go. I’ve got several 8 core systems that run a pair of Linux guests, and they only eat ~70% of the CPU cycles.

        • Flying Fox
        • 11 years ago

        Not when there’s a 8800GTX in the system. A GPU2 client needs to be there in order to maximize points.

          • JPinTO
          • 11 years ago

          I’d still run 4 Linux Vmware SMP clients and a GPU2 to max out the PPD… assuming that you have the RAM and disk to spare.

            • Gerbil Jedidiah
            • 11 years ago

            oops wrong spot

    • TheEmrys
    • 11 years ago

    It depends on OS, really.

    With Vista, you can generally use 2 SMP clients and a GPU client. However, in XP, the GPU client Utlizies a complete core.

    If you have Vista, use 2 smp clients (I’d strongly recommend 2 VM’s for them) and a GPU client. The ATI gpu client does not perform very well, but it does okay.

    If you have XP, I’d use 2 SMP. Your ppd will be higher than using the GPU.

      • bthylafh
      • 11 years ago

      Have they changed the SMP client to not need 4 cores for optimal speed, then? ISTR that’s how it was when I quit folding in May to save on cooling.

      • Flying Fox
      • 11 years ago

      That may not be so clear-cut if you are willing to play with affinity. So on XP 2xSMP on VMs, if you can have to -[

    • pikaporeon
    • 11 years ago

    … i swear I wont post again after this

    Stopping in the middle of the day is a BIT of a pain for your SMP clients, as their deadlines are short. that said you can still get them in. But because of that GPU folding is more tempting, as units finish in a few hours [but are only worth a few hundred points]

      • Flying Fox
      • 11 years ago

      The SMP WUs still have tight deadlines, so completely stopping them may not be such a hot idea. On my X2 box, I switched the vmware-vmx.exe to idle priority and it seems to be ok as long as I don’t run out of memory because of Firefox leaks. 😛

      WinSMP is ok if you are sure you won’t lose network connection (like you unplug the cable or reboot your router), or you will have to restart the client. However both WinSMP and GPU2 properly run their processes in Idle priority so for the most part it is not that intrusive, as long as you have enough RAM to avoid paging with your other running processes.

      Some gerbils seem to be ok even running 2xWinSMP on the system and are fine. I don’t know if you can do that with a GPU2 client. May be 1 WinSMP and 1 GPU2, or add another single core client there.

      You can also think about a VM that takes up 2 cores to run Linux SMP, a single core client and then a GPU2 client.

    • Krogoth
    • 11 years ago

    SMP Client is fast, but high-maintenance and not recommend for farms. Dr. Fish can attest to this.

    GPU client is a mixed bag. It does consume some CPU power, but depending on the GPU in question it can easily outrun even a fully utilized QX9770 in productivity. However, GPU and CPU point scales are completely different. 😉

    PS3 is a good folding box, BR-player and has some decent games for all just $499.

    The standard client works, but it can be tricky to set them up properly on a multi-core rig.

      • drfish
      • 11 years ago

      No kidding… I’m pretty sure my farm could do over 25k PPD (it was doing 20k+ over a year ago when leor and I were battling it out, even if it took a couple hours a day of maintenance) but it is so much work that I just stopped trying. Now I’m down to levels I haven’t been at in years as my old school farm of Athlon XP boxes retired to our manufacturing floor slowly keep the fish farm from going completely belly up by running the ancient and rock solid 4.x client (maybe 5.x, I don’t remember anymore).

      That’s why my sig on the forums instructs people to wake me when Windows SMP doesn’t suck anymore, I’d come back in an instant. Instead here we are a year later, with pretty solid GPU clients, the PS3 taking over the world, and the same fundamentally flawed Windows SMP clients! Yikes, can you tell I’m still a little hot over this? I even let the ones that were working die because I got sick of tracking when beta’s expired to do reinstalls of newer clients that ONLY served to extend the expiration date…

      Just last week I shut down my main machine at home, it was a Q6600/8800GTX running SMP and GPU – making enough points to keep me in the top 20, but just not worth the electricity for the hassle – now I *gasp* actually have my computer go into sleep mode for the first time since I’ve owned my own PC…

      It’s super depressing…

      Damage, if you want the most points for the least hassle, run one GPU client and 3 of the old school clients on that box.

        • Usacomp2k3
        • 11 years ago

        I feel you man.

        • Firestarter
        • 11 years ago

        I hear you, I completely quit folding due to the hassle of the SMP client. I felt it was impossible to accept that there are critical flaws in it that have been there for over a year. For the spirit of science, I worked around those flaws time and again, so that the WU’s might actually finish and be uploaded.

        I know the SMP client doesn’t sound as sexy as a GPU or Cell client, but the way in which the Pande group seemed to excuse themselves from the problems of the SMP client made me quit the whole thing.

          • tfp
          • 11 years ago

          And there is the answer the best Folding client combo is install no Folding client at all.

            • Convert
            • 11 years ago

            Riiiight…. The post lacks both humor and logic so I am not sure where you were going with it.

            At any rate the GUI and console versions are fine. Those would be the optimal choice for a reliable setup.

            • tfp
            • 11 years ago

            No kidding? I just don’t get the amount of bitching people do about the client and that they are going to “quit” because they aren’t getting enough points. Now that’s science.

            That’s the real joke.

            No matter the client the science is good from what you are saying so the problem to me seems to be that the people on the team are more concerned with points then science. Maybe that’s why the team isn’t growing.

            It seems so odd people react so harshly any time F@H is brought up in the forums or on the front page. The only thing I can think of is the way it’s brought up turns people off.

            • Convert
            • 11 years ago

            I didn’t really see anyone claiming they are going to quit (though I am probably not paying close attention) but I know some have so your point remains.

            More points = more science being completed. Whatever you use to measure, it represents your contribution. While you cannot say that we are all in this for the strict purpose of science it still plays a big role for most of us. The GPU and SMP clients get more work done and use more system resources so they are reimbursed for that extra work with more points. Can you say everyone complaining is strictly upset about not getting the points? Perhaps some want to be able to contribute more with the hardware they have but instead are given crappy software. Good intentions or not, using crap software starts to grate.

            You and a couple of other posters have, through current and prior folding related posts, seemed to try and invalidate folding through various means. Intentionally or not, perhaps you are just curious, but the questions and points you bring are almost always pointed/loaded. Just because someone folds for points doesn’t mean they shouldn’t fold. In this life great things are sometimes done for selfish reasons and frankly I would rather people fold for points than not fold at all if it came down to that in such a unequivocal manner. Of course I would rather they did it because they wanted to help, though even Stanford knows this wouldn’t be the best implementation for a long term uptake of the project. Which is precisely why we have the point system that we do.

            I used to come home every night from work and dial my internet connection, wait while a 3-5mb folding result was uploaded, wait while a new WU was downloaded, disconnect, move the phone line over to another client and start the process over again. As if anyone needs a reminder, dial up is slow, needless to say I wasn’t getting much sleep, often dozing off in my chair halfway through only to have to finish the process. Otherwise an entire day would be wasted and I could run the risk of my results being too old for use. I mention t his because while some may bitch and complain and come off as only wanting points they went through similar hardships as me STRICTLY in the name of science. Back then my machines weren’t able to product much PPD, especially when I had a hard ceiling of 1 result per machine per day. I only know this because I have been following folding for a while and know some of the stuff other folders have gone through. Don’t take all comments at face value I guess is what I am saying.

            Now that the points are so insane, rising to levels never seen before, perhaps some of us are loosing our heads and need a reminder about what this is all about. My first post on this article was simply trying to bring new people down to earth, to be happy with running the standard client because you are at least contributing. Getting caught up with the high output clients can only lead to disappointment if you don’t have tons of free time and I wanted to try and avoid that by sharing my experiences.

            • derFunkenstein
            • 11 years ago

            Uh, in the first several words of Firestarter’s comment, he said he completely quit. So you weren’t reading very closely.

            • Convert
            • 11 years ago

            I said saying not said 🙂

            Well actually I said “claiming they are going to quit”, but as I went on to say we both know people have quit over this type of thing.

            • tfp
            • 11 years ago

            I’m not sure how to respond. In some ways it’s seem like I or “people like me” have personally offended you when it comes to folding. I haven’t gone out of my way to “to try and invalidate folding”. I did get sick of the constant “but will it fold” posts and the seeming constant recruiting in unrelated threads. From other responses I have seen many other people were and still are tired of it as well. Something has to explain the backlash the project gets on the forum.

            I have pointed out problems in which part of the team complete crapped on until some people with common sense stepped in. I have also taken my time to in the past to help figure out what systems will do with for folding, though that was years ago now. I try not to even post in the F@H forum anymore and really haven’t much over the last few years. With a number of people there its like trying to talk to any zealot and just isn’t worth it most of the time, even if there is something useful to be added that could actually help.

            As for your first paragraph I’m sure I could find a few people who quit or at least greatly scale back because of points issue. However it isn’t worth the search and we both know it has happened.

            Paragraph 2, more points isn’t more science they weighing of points are completely skewed. I know why its done, so people will use the new apps but just because someone does 1 unit and it gives 1.5K points, it doesn’t mean that there has been more science done then 6 units at 250 points a piece. Because I’m not on the research team I can’t really say what is really more useful. My guess it really depends on what set they are looking at right now.

            Complaining about point totals and other items I have seen are not going through the same thing you are describing. I don’t see what your point is, if you are talking about the large memory issues of a couple years ago that are different. It is a true problem that did impact users. You do your self a disservice by grouping what you did to fold with those items.

            Perhaps people are loosing their heads about points again, and that is the cause of the push that will help science. However I haven’t seen most of the items in the “how to improve F@H thread at TR” implement. That includes things like time consuming stuff like website design and weekly write ups, simpler things like like cleaning up the sticks so they actually have useful info in them. What I do see is after a long periods of time there is a lot of talk which ends up being what others need to do for the team because of lack of points. It comes across in the same manor as people at work that sit in meeting all day saying what should be done (and they may or may not be right) but never actually do any real work.

            Because of “but will it fold” thread crapping, constant recruiting, show boating, pig headedness of some in the past and the stuff I have seen lately a long with the way the project is run and the zealotry at the folding@home forum I wouldn’t even consider joining and they are the same reasons I quit years ago. Looking at the Stanford site I had almost 400k points and other then a slow box an inlaw’s that might be submitting off and on I haven’t folded in ~3 years now, if I am remembering right. There’s probably +800k points worth of science not done, whatever that means.

            If people want to grow the team, to me at least, things would need to change. You might look at this as one big attempt to dis F@H, however I have posted this as an example of why someone quit. Maybe others have for some of the same reason, and a few changes might bring them back. Maybe not but points are at stake! 😉

            • Convert
            • 11 years ago

            q[

            • tfp
            • 11 years ago

            Ah that is somewhat related to same argument Vijay had in the past with running 2 WU on a P4 with HT and getting 20-30% more work done in the same time or turning in one results quicker. He was for turning in the one results quicker even because it “got more work done” and allowed them to take the result and generate the next work unit faster. To me it was flawed logic. The project still supported slower P3s and normal athlons at the time (maybe even P2s were fast enough back then), accepts people not running 24/7, and allowed the use of slower hardware. What’s the point of getting one work unit done a little faster if it is going to be assigned to a slow box when you can get two done with just a bit more time.

            The same thing would apply for the multi core units. Even if the scaling of the crunching was 100% per core, and its not, the same amount of calculations is being done if you have 4 cores working on one unit or 4 cores working on 4 units. The difference is more work is being done on a project at a time. So that project has more of a concentration of processing power then one that is part of the old client. With the short return times it allows them to push projects of importance, because they know only the hardware that is fast will be doing the work. Chances are it will have more ram as well. So even though the same amount of Flops are being done (just to use a measurement), because you need a system will a lot of ram, 4 cores, ect and deal with an unstable client they give bonus points. It is a way to direct processing power to projects they want worked on. At least in my eyes, which is makes sense from Stanford’s point of view. It’s the easiest way to redirect the powerful boxes to projects that are more important assuming they use the newer more resource intensive clients. Point values have always been a point of contention for different projects at F@H, because HW changes so much vs the base machine they use only so much can be done for every machine that works on the same WU. However the point inflation they do to me isn’t that great, but in some ways I can understand why they do it. However if points are your thing, looking back it makes it seem like work done in the past wasn’t as important. As a side note I think this is one of the main reasons they never released a BOINC client.

            Yep I think Stanford needs to do a better job with its clients, the problem is I don’t expect most of the people working on the projects have a programming background. Stanford really needs to work on the easy of use and upgrade ability of the project; I think BOINC would have help a lot in that regard. However, to me at least, that has never been a priority of theirs.

            In some ways that is understandable, in others it isn’t. I know some colleges have formed an “alliance” between the bio and cs department when it comes to bio-informatics type projects. I have actually worked with one of these groups and is part of my annoyance with all of this, the kind of ivory tower type thinking. Maybe Stanford has gotten better about that over the past few years but it doesn’t sound like it. They only have the resources that they have. Students are always the best programmers even if they have the theory down.

            The one thing that is brought up all the time is “has Stanford done anything with the data”. A quick check on the web site the other day showed they have 56 or 57 papers write on the finders from this folding project. That is at least a counter I don’t remember seeing to show the nay Sayers like my self. 😉 If anyone can actually dig though some of that stuff or even make an intelligent write up it could be an interesting discussion.

            Really to me simple things need to be done first, from there who knows. Lowering the annoyance level of hearing about the project on TR should be the first goal, unless I’m the only person annoyed. I haven’t read other feed back like mine so maybe it is just me.

            • Convert
            • 11 years ago

            I was under the impression that the SMP client is more efficient than running more than one console version and is able to produce more because of that. As in the true mflops of the processor doesn’t change, it’s just able to take better advantage of the hardware and use more of the processing power available to it and get more work done. Perhaps instead it is only a platform to prioritize important WU’s to higher end systems. Maybe both?

            As for the GPU client, well I guess that speaks for itself. While the points are still skewed it definitely has more of an excuse than the SMP client when it comes to reimbursement.

            • farmpuma
            • 11 years ago

            tfp wrote, “The same thing would apply for the multi core units. Even if the scaling of the crunching was 100% per core, and its not,”

            Wrong. All my cores, be they dual or quad run at 100% for 100% of the crunching time.

            Come on, admit it, you really want one of those G92 based video cards and to jump back into folding. 112 or 128 processing elements do a project GOOD.

            • Convert
            • 11 years ago

            I am not sure if that is what he was getting at. Even if you run four instances of the console client it also shows 100%, though of course both clients make room for other processes to use system resources.

            • tfp
            • 11 years ago

            My point was the SMP clients would not get more science done then the standard non-SMP clients assuming the same number of non-SMP optimizations are in the code.

            The only thing I can think of leading to that is if the memory bandwidth is much more of a bottle neck on running multiple single core clients vs. one multi threaded client. However to keep both cores feed it would hit the FSB x times more because more clients are running, then there is cache coherency over the FSB to contend with.

            In the past and now Stanford has always given more credit to resource intensive apps and has used point bonuses to get people to move to them. The resent performance improvement in the Multi-threaded apps further shows it is not getting all of the performance out of all of the cores it is using and would could argue that as an example of how multi core apps are probably not as efficient as multiple signal core apps.

            That is why I said points don’t mean more science. I guess the end statement should have been people are doing just as much science running 4 single core clients as someone running 1 multi threaded client even if they are not getting as many points. I’m not saying people should just run single core apps because of science, just that multi-threaded apps are kind of like the request to run one app on a P4 with HT. Stanford would rather get something done faster then get more things done a little slower. If they have important research that is kind of time dependent it makes a lot more sense with the multi threaded app because only fast machines can do it.

            Of course the only way to really know is if they produced some way to really measure the output using Flops or something of that nature. But from other apps people have benchmarked in the past, including the science apps TR benchmarks nothing scales 100%.

            GPU and PS3 can’t be directly compared because of different “processor” architectures.

            Does that make more sense?

            • tfp
            • 11 years ago

            Just because all 4 cores are being used doesn’t mean that some amount of performance is lost. I’m not sure anything scales 100% per core added, it can be close but I don’t believe 100% is possible. If it is using 4 cores and you are a core2 I know that cache coherence is going to impact because it needs to use the FSB to wait on that. I believe in those kinds of cases the CPU is pegged in task manager and waiting, same goes for FSB waiting, ect. One thing is for sure it will never scale more then 100%. So best case is 100% per core, so your doing the same amount of work as 4 singles assuming the same calculation optimizations not related to MP are implemented.

            • tfp
            • 11 years ago

            Here is my put up or shut up thread.
            F@H Info Thread
            §[<http://www.techreport.com/forums/viewtopic.php?f=9&t=61091<]§ In my posts out here I have stated people complain but do nothing about it. I can't to much but I can do what was listed in the thread above and I don't mind trying to keep it up to date. Now that I am part of the solution I can complain even more! §[<http://en.wikipedia.org/wiki/Festivus#Airing_of_Grievances<]§

          • just brew it!
          • 11 years ago

          The single-core clients are still “fire and forget”. It’s a tradeoff… if you’re willing to put in a little extra effort to babysit the Windows SMP client, you’re rewarded with extra points.

          The critical flaws of which you speak are due to the multi-threading framework they’re using to communicate between the threads running on each core. The MPI framework was designed for *NIX systems, and the Windows port of it is a bit buggy. Unfortunately, they are molecular biologists not computer scientists; furthermore they have limited resources. So yeah, it has taken a while for progress to be made on improving the stability of the Windows SMP client. But they do clearly state that it is beta software, so it is not like people who decide to try it haven’t been warned — it just isn’t for everyone.

            • derFunkenstein
            • 11 years ago

            But the fact that the Pande group hasn’t gotten a stable SMP client tells me that F@H isn’t a priority to them. Getting WU’s done as quickly as possible and having as many results as possible are what are going to help. If they don’t get good, stable clients that don’t crash often, they’re not only hurting themselves; they’re screwing with the people kind enough to donate spare CPU cycles (such as yourself).

    • pikaporeon
    • 11 years ago

    Oh, also when I tried two SMP clients and a GPU, the GPU went to a standstill with them all shjaring the work, and when I set two SMP to share 3 cores, it took over an hour in each SMP client per percent [this is on a Q9300@3.2]

    • pikaporeon
    • 11 years ago

    the 3850 is not a great folder. I’d set up two VM environments in VMWare, and run an SMP client in each under linux, or use notfred’s Diskless Folder VM image. Each client should turn over at least a 1700+ point WU a day.

    Dual core would be just one SMP client

    The Nvidia 8x series are the best folders right now (some hitting 5000 PPD), and a Radeon -4-850 produces a few PPD less than a single SMP client in a VM (around 1400).

      • Flying Fox
      • 11 years ago

      I would still say you should do a test to see if the 3850 is worth it.

      To over simplify things a bit, let’s say you do 1500ppd per SMP client if you run 2x SMP (3000ppd) on the quad. If the 3850 can do 3000ppd, then even if the SMP clients drop to 500ppd (2×500=1000ppd total) you are still ahead (3000+1000 =4000ppd).

        • pikaporeon
        • 11 years ago

        No ATI client even approaches 3000 PPD.

          • Flying Fox
          • 11 years ago

          Then what are they doing these days?

          My HD3450 is doing ~300ppd.

        • SmokinJoe-Salem
        • 11 years ago

        My setup: Q6600, Vista-32, HD4850, VMware-2 x Linux-64 All stock speeds, VM priorities lowered per Trax
        The two Linux clients were both running project 2662 with the a2 version 2.00 core and cranking out 1% increments in 14 min (+/1 a few seconds) for a combined calculated total of 3916 ppd. The GPU was fired up and let run overnight. The Linux clients dropped to 18.5 and 24.5 min and the gpu was cranking out 2:30 increments, down for 2.02 when running alone. Total output was 3769 ppd. The gpu client was counterproductive, a waste of electricity. Right now 1 SMP client + GPU is about a wash with the 2 SMP clients.

      • PRIME1
      • 11 years ago

      Surely their must be a spare GTX280 lying around the labs?

Pin It on Pinterest

Share This