afaik the only major win/lin difference at present is on amd. The athlon's choking on the compilers code for the hotloop and not optimizing it very well. Using the same compiler for the *nix app under windows didn't help any.
afaik the only major win/lin difference at present is on amd. The athlon's choking on the compilers code for the hotloop and not optimizing it very well. Using the same compiler for the *nix app under windows didn't help any.
I share Gary's suspicion that the performance figure for Annika's Core Duo under Windows is sub-optimal and that there's a gap between Windows and Linux for other platforms than AMD as well. I'm running a CD under Darwin which will produces > 15 cr / hr, and a Pentium M 1500 (which is very similar to a CD) under Linux which is outperforming Annika's CD in terms of cr/hr.
But: are we sure that the mechanism used under Windows and Linux to count CPU seconds used by an application is comparable and equally precise? Statistics, you know..... Gary, it would be also interesting to see what the difference in REAL time spent per WU is in your WIn vs Linux experiment, as I guess the hosts are otherwise unloaded.
... are we sure that the mechanism used under Windows and Linux to count CPU seconds used by an application is comparable and equally precise?
Yes, I believe so, if we are talking about an NT based Windows. I wouldn't trust Win9x ...
Quote:
... it would be also interesting to see what the difference in REAL time spent per WU is in your WIn vs Linux experiment, as I guess the hosts are otherwise unloaded.
Some time ago I went to the trouble of comparing (for WinXP) the reported CPU seconds with the elapsed wall clock seconds for a number of work units and found that there was minimal difference between the two - for an otherwise unloaded system. I seem to remember the wall clock time being a minute or two longer than the CPU time for a result taking about 5-6 hours. I have no reason to suspect that Linux would be any different. It's nearly 1.00am here now so tomorrow sometime I'll go through the log files on my new Linux machines and work out the elapsed wall clock time for the first completed results (which still have about an hour to completion, on the first machine. Results on the second machine have 2 full days left as I've just finished loading Linux and installing BOINC and the crunching has just started.
So as not to keep hijacking this thread, I'll start a new thread with the results of my experiment. So far it seems that crunching on Linux for a coppermine PIII gives about a 20-30% speedup as compared to WinXP - unless I've stuffed up the calculations somewhere ...
So far it seems that crunching on Linux for a coppermine PIII gives about a 20-30% speedup as compared to WinXP - unless I've stuffed up the calculations somewhere ...
This would be an amazing difference. I don't expect the OS itself to have any noticeable effect, it would be the compiler and maybe math library that causes the difference, right?
But you are right, we should continue in a separate thread about OS/Compiler related performance analysis.
Annika, any news about your Win vs Linux experiment ?
Unfortunately the only Win XP installation I have is too mangled with performance eating AV scanners, personal firewalls, WLAN drivers and other service like stuff to be comparable to a Linux installation on the same host.
Annika, any news about your Win vs Linux experiment ?
Unfortunately, I had to take "dual boot" a bit literally recently (meaning I used both OSs over the past few days) so the WU in question is not quite finished yet. I can give a status update, though: At exactly 50% finished, the WU has taken about 15 hours, meaning my initial estimate of 32 hours doesn't seem to have been too low, if all, it was a bit generous...
Annika, any news about your Win vs Linux experiment ?
Unfortunately, I had to take "dual boot" a bit literally recently (meaning I used both OSs over the past few days) so the WU in question is not quite finished yet. I can give a status update, though: At exactly 50% finished, the WU has taken about 15 hours, meaning my initial estimate of 32 hours doesn't seem to have been too low, if all, it was a bit generous...
That looks great for a 443 credits WU, much better than what you got previously under Windows. So what has previously been described as an AMD/Windows penalty might be a broader phenomenon. According to boincstats.com, more than 80% of the hosts are Windows based at Einstein@Home, so if the Windows code could be made as efficient as the Linux code, the total performance would benefit quite a lot.
BTW, a happy holiday & extended weekend to all the lucky guys & gals who could take Friday off (tomorrow is a holiday in Germany). Yup, this includes me *joy*
Yep, certainly :-D then they would have nothing to exploit cheaply and nothing to bash Bill Gates for. I think some of them might actually shoot or hang themselves, having lost meaning and purpose in life ;-) Btw, I like Linux myself, because it's Open-Source, safer than Windows and gives more configuration options. But I don't like it if people get too zealous about stuff like this. Why not let everyone use what suits him/her best? Okay, I don't think that Windows and servers go well together. Usage in computer centers or companies is imo for Linux/Unix only; the rest is more trouble than it's worth. But why freak out about a desktop system?
@Bikeman: Congratz to you :-) . Unfortunately, I couldn't get Friday off. Got another one of those annoying "practical lessons" then (not that I don't like working with computers but the profs tend to be really demotivating about it) so I have to show up at Uni.
To add something which is not completely off topic: at a good 70% crunched, the 30-hour-estimate still seems to be fairly exact. The other core is having fun with some little snacks of around 150 credits.
Okay, the WU in question just finished and uploaded. It shows exactly 29 hours, 54 minutes and 11 seconds on my CPU time counter and earned me 443.13 credits (yes, I was a bad girl and forced an immediate upload just once...), slightly outperforming a Pentium D 3.4 GHz running XP which benchmarks with a significantly higher floating point speed than my box (1750 compared to about 1400 on my box, both measured under Win XP... yes I know, my current benchmark looks lower but I had instant messaging and tons of browser tabs on at that time, and running mainly Einstein I didn't bother to re-run the benchmark) so it looks like the OS does make a difference after all... And my Windows installation is fairly new, plus I don't run an on-access virus scanner... deactivated some unnecessary Windows services, too, so there are no real performance killers around other than the stuff I also use under Linux (instant messaging, web browser, media player and stuff like that).
Some numbers: the current WU gave me roughly 14,8 credits/hour... 50 hours for 530 credits would be 10.6 credits per hour... so Windows has only 71.6% the performance of Linux on my box atm. Just figure out what ~30% more performance would cost if you went to buy a new CPU, and Linux is free ;-) All this bases on the idea that the credits are distributed reasonably fairly between different WU sizes, but I would be surprised if they weren't. Don't feel like making a whole survey about it, but it can't possibly account for 30% credit difference or someone would have bashed the staff about it long since. Interesting that this doesn't seem to be "an AMD problem" after all, at least not exclusively.
EDIT: Just asked my brother about CPU prices (he is normally a bit better informed about the market situation) and he estimated about 300 Euros for a comparable performance increase with German hardware prices... might differ a bit where you live but I think you get the picture.
Finally, I did install BOINC & E@H on the notebook under XP Home that usually runs E@H under Linux (Knoppix).
Surprisingly, on my particular box, there's no difference whatsoever so far between Linux and Win XP performance in terms of credits/CPU hr.
It's a Pentium M which (in the development history of Intel CPUs), is kind of similar to both it's architectural predecessor (Pentium III, where Gary found a significant performance delta) and it's successor (Core, as in your host).
Isn't that strange? And it's not that credits/hour is a bad measure, for every host I've got this metric is quite constant, even for WUs from different frequency timebands (="sizes")!
Anyway, we must be careful not to blame the operating system itself, as different compilers were used to build the app for the different OSs.
I have no idea what causes this, but it's not as simple as "AMD under Windows is slow for E@H", that's for sure.
So, will you continue to run that host under Linux? 30% is quite a difference.
afaik the only major win/lin
)
afaik the only major win/lin difference at present is on amd. The athlon's choking on the compilers code for the hotloop and not optimizing it very well. Using the same compiler for the *nix app under windows didn't help any.
RE: afaik the only major
)
I share Gary's suspicion that the performance figure for Annika's Core Duo under Windows is sub-optimal and that there's a gap between Windows and Linux for other platforms than AMD as well. I'm running a CD under Darwin which will produces > 15 cr / hr, and a Pentium M 1500 (which is very similar to a CD) under Linux which is outperforming Annika's CD in terms of cr/hr.
But: are we sure that the mechanism used under Windows and Linux to count CPU seconds used by an application is comparable and equally precise? Statistics, you know..... Gary, it would be also interesting to see what the difference in REAL time spent per WU is in your WIn vs Linux experiment, as I guess the hosts are otherwise unloaded.
CU
BRM
RE: ... are we sure that
)
Yes, I believe so, if we are talking about an NT based Windows. I wouldn't trust Win9x ...
Some time ago I went to the trouble of comparing (for WinXP) the reported CPU seconds with the elapsed wall clock seconds for a number of work units and found that there was minimal difference between the two - for an otherwise unloaded system. I seem to remember the wall clock time being a minute or two longer than the CPU time for a result taking about 5-6 hours. I have no reason to suspect that Linux would be any different. It's nearly 1.00am here now so tomorrow sometime I'll go through the log files on my new Linux machines and work out the elapsed wall clock time for the first completed results (which still have about an hour to completion, on the first machine. Results on the second machine have 2 full days left as I've just finished loading Linux and installing BOINC and the crunching has just started.
So as not to keep hijacking this thread, I'll start a new thread with the results of my experiment. So far it seems that crunching on Linux for a coppermine PIII gives about a 20-30% speedup as compared to WinXP - unless I've stuffed up the calculations somewhere ...
Cheers,
Gary.
RE: So far it seems that
)
This would be an amazing difference. I don't expect the OS itself to have any noticeable effect, it would be the compiler and maybe math library that causes the difference, right?
But you are right, we should continue in a separate thread about OS/Compiler related performance analysis.
Annika, any news about your Win vs Linux experiment ?
Unfortunately the only Win XP installation I have is too mangled with performance eating AV scanners, personal firewalls, WLAN drivers and other service like stuff to be comparable to a Linux installation on the same host.
CU
BRM
RE: Annika, any news about
)
Unfortunately, I had to take "dual boot" a bit literally recently (meaning I used both OSs over the past few days) so the WU in question is not quite finished yet. I can give a status update, though: At exactly 50% finished, the WU has taken about 15 hours, meaning my initial estimate of 32 hours doesn't seem to have been too low, if all, it was a bit generous...
RE: RE: Annika, any news
)
That looks great for a 443 credits WU, much better than what you got previously under Windows. So what has previously been described as an AMD/Windows penalty might be a broader phenomenon. According to boincstats.com, more than 80% of the hosts are Windows based at Einstein@Home, so if the Windows code could be made as efficient as the Linux code, the total performance would benefit quite a lot.
BTW, a happy holiday & extended weekend to all the lucky guys & gals who could take Friday off (tomorrow is a holiday in Germany). Yup, this includes me *joy*
CU
BRM
RE: so if the Windows code
)
I know I'm taking this out of context, but boy, wouldn't there be a whole bunch of 1337 h4x0rz out there who would be mad if that was the case? ;)
Yep, certainly :-D then they
)
Yep, certainly :-D then they would have nothing to exploit cheaply and nothing to bash Bill Gates for. I think some of them might actually shoot or hang themselves, having lost meaning and purpose in life ;-) Btw, I like Linux myself, because it's Open-Source, safer than Windows and gives more configuration options. But I don't like it if people get too zealous about stuff like this. Why not let everyone use what suits him/her best? Okay, I don't think that Windows and servers go well together. Usage in computer centers or companies is imo for Linux/Unix only; the rest is more trouble than it's worth. But why freak out about a desktop system?
@Bikeman: Congratz to you :-) . Unfortunately, I couldn't get Friday off. Got another one of those annoying "practical lessons" then (not that I don't like working with computers but the profs tend to be really demotivating about it) so I have to show up at Uni.
To add something which is not completely off topic: at a good 70% crunched, the 30-hour-estimate still seems to be fairly exact. The other core is having fun with some little snacks of around 150 credits.
Okay, the WU in question just
)
Okay, the WU in question just finished and uploaded. It shows exactly 29 hours, 54 minutes and 11 seconds on my CPU time counter and earned me 443.13 credits (yes, I was a bad girl and forced an immediate upload just once...), slightly outperforming a Pentium D 3.4 GHz running XP which benchmarks with a significantly higher floating point speed than my box (1750 compared to about 1400 on my box, both measured under Win XP... yes I know, my current benchmark looks lower but I had instant messaging and tons of browser tabs on at that time, and running mainly Einstein I didn't bother to re-run the benchmark) so it looks like the OS does make a difference after all... And my Windows installation is fairly new, plus I don't run an on-access virus scanner... deactivated some unnecessary Windows services, too, so there are no real performance killers around other than the stuff I also use under Linux (instant messaging, web browser, media player and stuff like that).
Some numbers: the current WU gave me roughly 14,8 credits/hour... 50 hours for 530 credits would be 10.6 credits per hour... so Windows has only 71.6% the performance of Linux on my box atm. Just figure out what ~30% more performance would cost if you went to buy a new CPU, and Linux is free ;-) All this bases on the idea that the credits are distributed reasonably fairly between different WU sizes, but I would be surprised if they weren't. Don't feel like making a whole survey about it, but it can't possibly account for 30% credit difference or someone would have bashed the staff about it long since. Interesting that this doesn't seem to be "an AMD problem" after all, at least not exclusively.
EDIT: Just asked my brother about CPU prices (he is normally a bit better informed about the market situation) and he estimated about 300 Euros for a comparable performance increase with German hardware prices... might differ a bit where you live but I think you get the picture.
@Annika Interesting
)
@Annika
Interesting stuff!!
Finally, I did install BOINC & E@H on the notebook under XP Home that usually runs E@H under Linux (Knoppix).
Surprisingly, on my particular box, there's no difference whatsoever so far between Linux and Win XP performance in terms of credits/CPU hr.
It's a Pentium M which (in the development history of Intel CPUs), is kind of similar to both it's architectural predecessor (Pentium III, where Gary found a significant performance delta) and it's successor (Core, as in your host).
Isn't that strange? And it's not that credits/hour is a bad measure, for every host I've got this metric is quite constant, even for WUs from different frequency timebands (="sizes")!
Anyway, we must be careful not to blame the operating system itself, as different compilers were used to build the app for the different OSs.
I have no idea what causes this, but it's not as simple as "AMD under Windows is slow for E@H", that's for sure.
So, will you continue to run that host under Linux? 30% is quite a difference.
CU
BRM