Message boards :
Number crunching :
Windows vs. Linux
Message board moderation
Previous · 1 · 2 · 3 · 4 · 5 · Next
Author | Message |
---|---|
Josef W. Segur Send message Joined: 30 Oct 99 Posts: 4504 Credit: 1,414,761 RAC: 0 |
... The stock 5.27 Windows build was made with DevC++/MinGW which uses the GCC 3.4.2 toolchain. The stock 5.28 Linux build identifies itself as being produced by GCC 4.1.2. There are multiple differences between those versions of GCC, and it's not hard to find statements on the internet that GCC 2.x or 3.x or 4.x produces the fastest code. But I don't think the specific cases cited involve the large percentage differences DeMus has observed. The other factor is what must be done to be compatible with the various different versions of Windows vs. what is required to support the full range of Linux distros. As Crunch3r pointed out earlier, using the same source does not mean the builds are identical; all modern compilers attempt to optimize their output though some are more successful than others. There are a lot of choices in options for the compilers, too. Whoever makes the builds can control what the compiler does to a large extent, but there are so many options that it would take a long time to carefully choose the best set. The Windows and Linux stock builds are related, but it's a mistake to think of them as identical twins; if both Operating Systems were capable of running the same app the project wouldn't be producing separate apps. Joe |
Steven - KO4E Send message Joined: 21 Jun 99 Posts: 53 Credit: 2,434,487 RAC: 0 |
Hi. Remember me? ;-) systeminfo is in XP media addition. Wow a whole lot of info there. A few weeks ago I tried ubuntu 64bit with crunch3r 64bit app and ran it for 2 or 3 days. This was on a core2duo 6300 dell e520 without graphics. I every result I checked it was faster than stock windows or macs of comparable speed. It even was faster then q6600 quads with stock apps. I did not come up against any optimized clients :( It seemed to really scream on the longer work units. I would have stayed with it but I need to run XP for now on the machine :( SETI@home classic workunits 5,429 SETI@home classic CPU time 73,472 hours |
ML1 Send message Joined: 25 Nov 01 Posts: 21212 Credit: 7,508,002 RAC: 20 |
Hi, it's me again. Remember me: the iniator of this thread who was trying to find out if there are more people having similar results as I have in using the two above mentioned systems. Don't worry, no blame there. You might say that old battles die hard. Silly really because on this subject there shouldn't be a 'battle'. But also, there often seems to be great sensitivity or zealotry until whatever 'light' is seen. I found out something on my computer and was trying to see if other users have seen it to, or maybe just the opposite. And that is very interesting and a curiously large difference. More information is needed to be able to see what is happening. Better still would be if someone else can replicate the same effect to then do some investigating and debug. Knowing why will be useful. As linked earlier, you can also have completely the reverse where Linux appears faster than Windows. In the ideal, they should both be within just a few percent of each other. Thanks Jo for the compiler comments. This thread is realy going the wrong way, maybe I should have known better and shouldn't have started this thread. My apologies for having done so. No at all. However, it's a little disappointing when the beginnings of a good discusion descends in to glib silly 'sound-bites'. I have returned to using Windows for crunching since it gives me 50% extra power compared to my installation of Linux, not because I want to. I wanted to get some experiences with Linux and in the mean time do the crunching. Take a look at WINE on Linux. The WINE 'utility' lets you run Windows applications on Linux. It is not an emulator so you should get the full Windows speed when using it. Check if WINE is installed and then install the Windows Boinc in there. See if it works. I'm sure others will have done that. It's exactly what I did to get maximum speed for s@h-classic on an old Linux install of mine. All long, long ago. I will keep on following this thread, for amusent only, since an answer to my question will be out of the question completely. Should be good for a giggle. Please keep asking questions! Happy crunchin', Martin See new freedom: Mageia Linux Take a look for yourself: Linux Format The Future is what We all make IT (GPLv3) |
Gerry Green Send message Joined: 15 May 99 Posts: 13 Credit: 3,034,952 RAC: 0 |
More information is needed to be able to see what is happening. Better still would be if someone else can replicate the same effect to then do some investigating and debug. Knowing why will be useful. I've posted a couple of times above in the thread -- I have a similar setup and I'm getting very similar results. Q6600 Quad, openSUSE 64 and Windows Vista 64 (to see stock client timings look for work units returned prior to 11 Mar -- I switched to the optimized client after that). The stock clients on each side were showing a very similar ratio as to what was reported by the original poster. 72 credits took about 11k on windows but 16k or so on linux. My benchmarks are also similar. I run no graphics or screen savers on either side. I've been running the optimized client on windows for a couple of days now and getting 7500 or so seconds for 72 credits. I'll boot over to the linux side sometime soon and try the optimized client over there. This maybe more curiosity than anything but the CPU fan speed would make it seem the windows BOINC client is keeping the CPU a lot hotter than it does on linux... On linux with just BOINC running I never hear the fan switch in to higher speed. I noticed this a couple of weeks ago while working with oracle -- when I'd launch a big query the fan would speed up. I even started using that effect to note when my queries were complete as the fan would go back to "normal". With or without a query running, top was reporting all CPUs at 100%. (note that I do not use Oracle all that often and it's installed on both sides of the machine so I doubt it changes this much...) Under windows BOINC alone will consistently drive the CPU fan speed to the same levels it did while the Oracle query was running under linux. |
Brian Silvers Send message Joined: 11 Jun 99 Posts: 1681 Credit: 492,052 RAC: 0 |
... So, in a nutshell, you're saying I'm right...that the Linux GCC build was "challenging" because the poor saps were undermined by the legion of evil? ;-P Edit: Yeah, I'm being facetious... If you want to expand further upon your explanation, please do, since it is what the thread is all about... :-) |
Brian Silvers Send message Joined: 11 Jun 99 Posts: 1681 Credit: 492,052 RAC: 0 |
But also, there often seems to be great sensitivity or zealotry until whatever 'light' is seen. I think it was George Carlin, but it might have been someone else...but they had this routine about how after a night of drinking people wake up late and stumble out and see the 'light', which is "God's Flashlight" (the sun)... ;-) In the ideal, they should both be within just a few percent of each other. Yes, I agree, and have said as much over at Einstein... I think I had stated a 3-5% variance was acceptable to me...but I'll freely admit that I'm a novice as far as compilations go...particularly with scientific applications. |
ML1 Send message Joined: 25 Nov 01 Posts: 21212 Credit: 7,508,002 RAC: 20 |
More information is needed to be able to see what is happening. Better still would be if someone else can replicate the same effect to then do some investigating and debug. Knowing why will be useful.I've posted a couple of times above in the thread -- I have a similar setup and I'm getting very similar results. Q6600 Quad, openSUSE 64 and Windows Vista 64 (to see stock client timings look for work units returned prior to 11 Mar -- I switched to the optimized client after that). Thanks for the useful comments and confirmation. This maybe more curiosity than anything but the CPU fan speed would make it seem the windows BOINC client is keeping the CPU a lot hotter than it does on linux... Now that is interesting... Does the Linux distro you are running have "power saving" of some sort enabled by default? What might be happening is that for low priority tasks, the CPU speed is being throttled back (aka 'speedstep' or 'cool-n-quiet'.)... A comparison between the optimised apps on your known system will be good also. Thanks again for bringing some useful comments to the thread. Happy crunchin', Martin See new freedom: Mageia Linux Take a look for yourself: Linux Format The Future is what We all make IT (GPLv3) |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14679 Credit: 200,643,578 RAC: 874 |
Now that is interesting... Does the Linux distro you are running have "power saving" of some sort enabled by default? What might be happening is that for low priority tasks, the CPU speed is being throttled back (aka 'speedstep' or 'cool-n-quiet'.)... So does that mean we've come full circle back to message 724103? Martin, would you happen to know which Linux distros have that CPU throttling for niced processes enabled by default? I've seen it mentioned a couple of times on the boards, but I can't remember the phraseology exactly enough to target a search on it. And you have a more encyclopædic knowledge of Linux that I do. The posts so far by DeMus and Gerry Green don't suggest that they (or their system manufacturers) have installed anything third-party or hardware related to achieve this effect - and I thought 'cool-n-quiet' was an AMD-specific technology: both exemplars are using Intel CPUs. |
Brian Silvers Send message Joined: 11 Jun 99 Posts: 1681 Credit: 492,052 RAC: 0 |
And you have a more encyclopædic knowledge of Linux that I do. Mine's more pedantic...
Speedstep is Intel though... Seriously though, the compilers used are different. It could be that sufficient runtime length testing was not done over on the beta project and so one platform is less tuned than the other...much like how the stalling was found over at Einstein, then worked around... @DeMus... I had to pick on Martin. It's an unwritten rule... ;-P I see him like "Ash" from the "Evil Dead" franchise, and Microsoft being the witch/undead/whatever that says "I'll swallow your soul". He responds, "Come get some." @Martin - no, this does not mean that I think you'll be "victorious" in the end... :-P |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14679 Credit: 200,643,578 RAC: 874 |
So, an hour and a half (and twelve miles) later, I still have to run my own search while fixing a client's machine. How does message 560625 sound? You might try removing powernowd with synaptics package manager. I had this problem when I first started using Ubuntu and this is how I got BOINC running full bore. Were either of you using Ubuntu? I've lost the will to look it up. Edit: And similar for Gnome - message 665824. You can do the rest yourselves. |
Gerry Green Send message Joined: 15 May 99 Posts: 13 Credit: 3,034,952 RAC: 0 |
So, an hour and a half (and twelve miles) later, I still have to run my own search while fixing a client's machine. How does message 560625 sound? Thanks for the info -- that is almost certainly what is going on... Sorry, I would've searched myself but 1. I wasn't actually trying to fix it, just helping out the original poster with another example, and 2. it really didn't seem like there was such an obvious answer after the thread had gone on this long and all the discussion of benchmark problems and compiler difference. Anyway, now that I know this is an option I'll still probably leave it like it currently is -- heckuva lot quieter. |
DeMus Send message Joined: 5 Jan 08 Posts: 238 Credit: 1,765,862 RAC: 0 |
Hi, A few things: I keep reading about an optimized client. What is it and where can I get it? I mean when I can do more in the same time, with the same computer then I also want to use the optimized version. I see/hear no differences in fan speeds between Linux and Windows. Both system remain cool and produce about the same soundlevel. On both systems I have a standard installation (Windows XP-64 and Fedora Core 8, 64 bit) with no settings done for powermanagement. Does this statement help? ______ DeMus |
Alinator Send message Joined: 19 Apr 05 Posts: 4178 Credit: 4,647,982 RAC: 0 |
Well your confusion about 'optimized clients' is most likely coming from a common mis-usage of the terminology in the BOINC context. In the Classic days the term applied since it was essentially one one big monolithic blob, with no distinction between the part which crunched the numbers and the part which did everything else. With BOINC you need to be clear about whether you're talking about the Science Application (which currently crunches the numbers and handles the 'screensaver' graphics) and the BOINC Core Client (which handles everything else). In your case I'm assuming you're interested in the optimized application, and there is a sticky thread here in NC for info about them. HTH, Alinator |
Gerry Green Send message Joined: 15 May 99 Posts: 13 Credit: 3,034,952 RAC: 0 |
Here's a link to the optimized client thread. I downloaded the ones from Crunch3r's website. I knew those were available previously but this thread piqued my interest enough to go find them. I'm not at home to check but my guess is the key to solving this is here. Basically sounds like the seti processes just have a lower priority (nice) and linux has a feature that won't push your processor to it's limit for threads of that nature. Knowing that feature is out there I may just leave it as is... Hard to say on your fan speeds -- that would be higly dependent on your hardware and other factors. I probably wouldn't have even noticed the difference if I hadn't stumbled on a way to trigger the faster speeds under linux... Based on everything else we've compared though I suspect it's highly likely we have the same issue. |
W-K 666 Send message Joined: 18 May 99 Posts: 19399 Credit: 40,757,560 RAC: 67 |
The Optimised apps are available at Crunch3r's site http://calbe.dw70.de/seti.html. Regards Win V Linux we would like to know times to crunch very similar units. That is units with the same angle_range (AR), they should be claiming the same credits +/- 0.01. The AR is displayed on the result page if you click on its Task ID. It's ok I found two units Windows (Microsoft Windows XP Professional x64 Edition , Service Pack 2) computer id=776378712 CPU time 10390.5 sec, WU true angle range is : 0.396046, 72.51cr. Linux (Ubuntu 4.1.2-0ubuntu4)id=779532694 CPU time 15551 sec, WU true angle range is : 0.396309, 72.51. Linux machine takes ~50% more time for same work. |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14679 Credit: 200,643,578 RAC: 874 |
Regards Win V Linux we would like to know times to crunch very similar units. That is units with the same angle_range (AR), they should be claiming the same credits +/- 0.01. The AR is displayed on the result page if you click on its Task ID. Those two look very similar to the two I dug out and asked Martin to opine on a long, long way down this thread. It's also the reason DeMus split this topic out of his previous thread. The question is, as it always has been : WHY? |
Gerry Green Send message Joined: 15 May 99 Posts: 13 Credit: 3,034,952 RAC: 0 |
Those two look very similar to the two I dug out and asked Martin to opine on a long, long way down this thread. It's also the reason DeMus split this topic out of his previous thread. The question is, as it always has been : WHY? Didn't you find previous threads answering that about 10 posts up? Basically by default linux is deliberately not pushing the processor as hard as it could for the BOINC threads since they are lower priority. Unfortunately I won't be able to check out the solutions you found until this evening but I'm pretty sure this is no longer a mystery... |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14679 Credit: 200,643,578 RAC: 874 |
Those two look very similar to the two I dug out and asked Martin to opine on a long, long way down this thread. It's also the reason DeMus split this topic out of his previous thread. The question is, as it always has been : WHY? I think I wrote previous threads answering that about 10 posts up - or down, in my preferred sort order. I was just trying to show Andy (WinterKnight) that he was leading us round in circles again. Perhaps I was trying to be too subtle. |
Gerry Green Send message Joined: 15 May 99 Posts: 13 Credit: 3,034,952 RAC: 0 |
Those two look very similar to the two I dug out and asked Martin to opine on a long, long way down this thread. It's also the reason DeMus split this topic out of his previous thread. The question is, as it always has been : WHY? ahh, okay. yes your subtlety escaped me :-). |
W-K 666 Send message Joined: 18 May 99 Posts: 19399 Credit: 40,757,560 RAC: 67 |
Ok, yes we are going round in circles. Therefore there is only one thing to do and that is to get all those with identical hardware running Widows and Linux or dual boot, to report to Eric and Josh at the SETI@home/AstroPulse Beta Site and crunch some units on both systems and make sure it doesn't happen on version 6 or on Astropulse (AP). Oh, one word of warning AP units on core2 E6600 or Q6600 take 80+ hrs. |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.